Social Computing, Behavioral-Cultural Modeling and Prediction

Lecture Notes in Computer Science Commenced Publication in 1973 Founding and Former Series Editors: Gerhard Goos, Juris Hartmanis, and Jan van Leeuwen

Editorial Board David Hutchison Lancaster University, UK Takeo Kanade Carnegie Mellon University, Pittsburgh, PA, USA Josef Kittler University of Surrey, Guildford, UK Jon M. Kleinberg Cornell University, Ithaca, NY, USA Alfred Kobsa University of California, Irvine, CA, USA Friedemann Mattern ETH Zurich, Switzerland John C. Mitchell Stanford University, CA, USA Moni Naor Weizmann Institute of Science, Rehovot, Israel Oscar Nierstrasz University of Bern, Switzerland C. Pandu Rangan Indian Institute of Technology, Madras, India Bernhard Steffen TU Dortmund University, Germany Madhu Sudan Microsoft Research, Cambridge, MA, USA Demetri Terzopoulos University of California, Los Angeles, CA, USA Doug Tygar University of California, Berkeley, CA, USA Gerhard Weikum Max Planck Institute for Informatics, Saarbruecken, Germany

6589

John Salerno Shanchieh Jay Yang Dana Nau Sun-Ki Chai (Eds.)

Social Computing, Behavioral-Cultural Modeling and Prediction 4th International Conference, SBP 2011 College Park, MD, USA, March 29-31, 2011 Proceedings

13

Volume Editors John Salerno AFRL/RIEF, 525 Brooks Road, Rome, NY 13441, USA E-mail: [email protected] Shanchieh Jay Yang Rochester Institute of Technology, Department of Computer Engineering 83 Lomb Memorial Drive, Bldg 09, Rochester, NY 14623-5603, USA E-mail: [email protected] Dana Nau University of Maryland, Department of Computer Science A. V. Williams Building, College Park, MD 20742, USA E-mail: [email protected] Sun-Ki Chai University of Hawaii, Department of Sociology 2424 Maile Way 247, Honolulu, HI 96822, USA E-mail: [email protected]

ISSN 0302-9743 e-ISSN 1611-3349 ISBN 978-3-642-19655-3 e-ISBN 978-3-642-19656-0 DOI 10.1007/978-3-642-19656-0 Springer Heidelberg Dordrecht London New York Library of Congress Control Number: 2011922188 CR Subject Classification (1998): H.3, H.2, H.4, K.4, J.3, H.5 LNCS Sublibrary: SL 3 – Information Systems and Application, incl. Internet/Web and HCI

© Springer-Verlag Berlin Heidelberg 2011 This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, re-use of illustrations, recitation, broadcasting, reproduction on microfilms or in any other way, and storage in data banks. Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its current version, and permission for use must always be obtained from Springer. Violations are liable to prosecution under the German Copyright Law. The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. Typesetting: Camera-ready by author, data conversion by Scientific Publishing Services, Chennai, India Printed on acid-free paper Springer is part of Springer Science+Business Media (www.springer.com)

Preface

Welcome to the 2011 international conference on Social Computing, Behavioral– Cultural Modeling and Prediction (SBP11). The overall goal of the conference is to bring together a diverse set of researchers/disciplines to promote interaction and assimilation so as to understand social computing and behavior modeling. In 2008, the first year of SBP, we held a workshop and had 34 papers submitted; in 2010 we had grown to a conference and had 78 submissions. This year, our fourth, we continued to expand, joining forces with the International Conference on Cultural Dynamics (and amending our name accordingly) and we received a record 88 submissions. The submissions were from Asia, Oceania, Europe, and the Americas. We are extremely delighted that the technical program encompassed keynote speeches, invited talks, and high-quality contributions from multiple disciplines. We truly hope that our collaborative, exploratory research can advance the emerging field of social computing. This year we also continued offering four pre-conference tutorials, a cross-disciplinary round table session, a poster board session and a series of panels including a panel featuring program staff from federal agencies discussing potential research opportunities. The accepted papers cover a wide range of interesting topics: (i) social network analysis such as social computation, complex networks, virtual organization, and information diffusion; (ii) modeling including cultural modeling, statistical modeling, predictive modeling, cognitive modeling, and validation process; (iii) machine learning and data mining, link prediction, Bayesian inference, information extraction, information aggregation, soft information, content analysis, tag recommendation, and Web monitoring; (iv) social behaviors such as largescale agent-based simulations, group interaction and collaboration, interventions, human terrain, altruism, violent intent, and emergent behavior; (v) public health such as alcohol abuse, disease networks, pandemic influenza, and extreme events; (vi) cultural aspects including cultural consensus, coherence, psycho-cultural situation awareness, cultural patterns and representation, population beliefs, evolutionary economics, biological rationality, perceived trustworthiness, and relative preferences; and (vii) effects and search, for example, temporal effects, geospatial effects, coordinated search, and stochastic search. It is interesting that if we were to trace these keywords back to the papers, we would find natural groups of authors of different papers attacking similar problems. While it may seem at first glance that the topics covered by the papers are too disparate to summarize in any succinct fashion, there are certain patterns that reflect trends in the field of social computing. One is the increasing participation of the human and social sciences in social computing, as well as the active collaboration between such fields and science and engineering fields. Disciplines represented at this conference include computer science, electrical

VI

Preface

engineering, psychology, economics, sociology, and public health. A number of interdisciplinary and applied research institutions are also represented. This conference cannot be run by only a few. We would like to first express our gratitude to all the authors for contributing an extensive range of research topics showcasing many interesting research activities and pressing issues. The regret is ours that due to the space limit, we could not include as many papers as we wished. We thank the Program Committee members for helping review and providing constructive comments and suggestions. Their objective reviews significantly improved the overall quality and content of the papers. We would like to thank our keynote and invited speakers for presenting their unique research and views. We deeply thank the members of the Organizing Committee for helping to run the workshop smoothly; from the call for papers, the website development and update, to proceedings production and registration. Last but not least, we sincerely appreciate the support from the University of Maryland Cultural Dynamics Lab along with the following federal agencies: Air Force Office of Scientific Research (AFOSR), Air Force Research Laboratory (AFRL), Office of Behavioral and Social Sciences Research and the National Institute on Drug Abuse at the National Institutes of Health, Office of Naval Research (ONR), Army Research Office (ARO) and the National Science Foundation (NSF). We would also like to thank Alfred Hofmann from Springer and Lisa Press from the University of Maryland. We thank all for their kind help, dedication and support that made SBP11 possible. March 2010

John Salerno Shanchieh Jay Yang Dana Nau Sun-Ki Chai

Organization

Conference Chairs Program Chairs Steering Committee Advisory Committee

Tutorial Chair Workshop Chairs Poster Session Chair Sponsorship Committee Student Arrangement Chair Publicity Chair Web Master

Dana Nau, Sun-Ki Chai John Salerno, Shanchieh Jay Yang Huan Liu, John Salerno, Sun-Ki Chai, Patricia Mabry, Dana Nau, VS Subrahmanian Rebecca Goolsby, Terrance Lyons, Patricia L. Mabry, Fahmida Chowdhury, Jeff Johnston Anna Nagurney Fahmida N. Chowdhury, Bethany Deeds Lei Yu Huan Liu, L. Patricia Mabry Patrick Roos Inon Zuckerman Mark Y. Goh, Peter Kovac, Kobi Hsu, Rayhan Hasan

Technical Review Committee Myriam Abramson, Naval Research Laboratory, USA Donald Adjeroh, Western Virginia University, USA Agarwal, University of Arkansas at Little Rock, USA Denise Anthony, Dartmouth College, USA Joe Antonik, AFRL (RI), USA Gurdal Arslan, University of Hawaii, USA Erik Augustson, NIH, USA Chitta Baral, Arizona State University, USA Joshua Behr, Old Dominion University, USA Geoffrey Barbier, AFRL (RH), USA Herbert Bell, AFRL (RH), USA Lashon Booker, The MITRE Corporation, USA Nathan Bos, John Hopkins University – APL, USA

Douglas Boulware, AFRL (RI), USA Jiangzhou Chen, Virginia Tech, USA Xueqi Cheng, CAS, P.R. China Alvin Chin, Nokia Research Center David Chin, University of Hawaii, USA Jai Choi, Boeing, USA Hasan Davulcu, Arizona State University, USA Brian Dennis, Lockheed Martin, USA Guozhu Dong, Wright State University, USA Richard Fedors, AFRL (RI), USA Laurie Fenstermacher, AFRL (RH), USA William Ferng, Boeing, USA Clay Fink, John Hopkins University – APL, USA Anthony Ford, AFRL (RI), USA Anna Haake, RIT, USA Jessica Halpin, Not Available

VIII

Organization

Walter Hill, Not Available Michael Hinman, AFRL (RI), USA Jang Hyun Kim, University of Hawaii at Manoa, USA Terresa Jackson, Not Available Ruben Juarez, University of Hawaii, USA Byeong-Ho Kang, University of Tasmania, Australia Douglas Kelly, AFRL (RH), USA Masahiro Kimura, Ryukoku University, Japan Irwin King, Chinese University of Hong Kong Minseok Kwon, RIT, USA Alexander H. Levis, George Mason University, USA Huan Liu, Arizona State University, USA Lustrek, Jozef Stefan Institute, Slovenia Patricia L. Mabry, National Institutes of Health, USA K. Clay McKell, Not Available Sai Motoru, Arizona State University, USA Hiroshi Motoda, Osaka University and AOARD, Japan Anna Nagurney, University of Mass Amherst, USA Keisuke Nakao, University of Hawaii at Hilo, USA Kouzou Ohara, Aoyama Gakuin University, Japan Esa Rantanen, RIT, USA Bonnie Riehl, AFRL (RH), USA

Kazumi Saito, University of Shizuoka, Japan Antonio Sanfilippo, Pacific Northwest National Laboratory, USA Hessam Sarjoughian, Arizona State University, USA Arunabha Sen, Arizona State University, USA Vincient Silenzio, Not Available Adam Stotz, CUBRC, USA Gary Strong, Johns Hopkins University, USA George Tadda, AFRL (RI), USA Lei Tang, Arizona State University, USA Shusaku Tsumoto, Shimane University, Japan Trevor VanMierlo, Not Available Changzhou Wang, Boeing, USA Zhijian Wang, Zhejiang University, China Rik Warren, AFRL (RH), USA Xintao Wu, University of North Carolina at Charlotte, USA Ronald Yager, Iona College, USA Laurence T. Yang, STFX, Canada Shanchieh Jay Yang, RIT, USA Michael Young, AFRL (RH), USA Lei Yu, Binghamton University, USA Philip Yu, University of Chicago, USA Laura Zavala, Not Available Daniel Zeng, University of Arizona, USA Jianping Zhang, The MITRE Corporation, USA Zhongfei Zhang, Not Available Inon Zuckerman, UMD

Table of Contents

Using Models to Inform Policy: Insights from Modeling the Complexities of Global Polio Eradication (Keynote) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Kimberly M. Thompson Toward Culturally Informed Option Awareness for Influence Operations with S-CAT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ken Murray, John Lowrance, Ken Sharpe, Doug Williams, Keith Grembam, Kim Holloman, Clarke Speed, and Robert Tynes Optimization-Based Influencing of Village Social Networks in a Counterinsurgency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Benjamin W.K. Hung, Stephan E. Kolitz, and Asuman Ozdaglar Identifying Health-Related Topics on Twitter: An Exploration of Tobacco-Related Tweets as a Test Topic . . . . . . . . . . . . . . . . . . . . . . . . . . . . Kyle W. Prier, Matthew S. Smith, Christophe Giraud-Carrier, and Carl L. Hanson Application of a Profile Similarity Methodology for Identifying Terrorist Groups That Use or Pursue CBRN Weapons . . . . . . . . . . . . . . . . . . . . . . . . Ronald L. Breiger, Gary A. Ackerman, Victor Asal, David Melamed, H. Brinton Milward, R. Karl Rethemeyer, and Eric Schoon Cognitive Aspects of Power in a Two-Level Game . . . . . . . . . . . . . . . . . . . . Ion Juvina, Christian Lebiere, Jolie Martin, and Cleotilde Gonzalez

1

2

10

18

26

34

Predicting Demonstrations’ Violence Level Using Qualitative Reasoning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Natalie Fridman, Tomer Zilberstein, and Gal A. Kaminka

42

Abstraction of an Affective-Cognitive Decision Making Model Based on Simulated Behaviour and Perception Chains . . . . . . . . . . . . . . . . . . . . . . . . . Alexei Sharpanskykh and Jan Treur

51

Medicare Auctions: A Case Study of Market Design in Washington, DC (Keynote) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Peter Cramton

60

A Longitudinal View of the Relationship between Social Marginalization and Obesity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Andrea Apolloni, Achla Marathe, and Zhengzheng Pan

61

X

Table of Contents

Open Source Software Development: Communities’ Impact on Public Good . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Helena Garriga, Sebastian Spaeth, and Georg von Krogh Location Privacy Protection on Social Networks . . . . . . . . . . . . . . . . . . . . . Justin Zhan and Xing Fang Agent-Based Models of Complex Dynamical Systems of Market Exchange (Keynote) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Herbert Gintis Computational and Statistical Models: A Comparison for Policy Modeling of Childhood Obesity (Panel) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Patricia L. Mabry, Ross Hammond, Terry T-K Huang, and Edward Hak-Sing Ip Detecting Changes in Opinion Value Distribution for Voter Model . . . . . . Kazumi Saito, Masahiro Kimura, Kouzou Ohara, and Hiroshi Motoda Dynamic Simulation of Community Crime and Crime-Reporting Behavior . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Michael A. Yonas, Jeffrey D. Borrebach, Jessica G. Burke, Shawn T. Brown, Katherine D. Philp, Donald S. Burke, and John J. Grefenstette Model Docking Using Knowledge-Level Analysis . . . . . . . . . . . . . . . . . . . . . Ethan Trewhitt, Elizabeth Whitaker, Erica Briscoe, and Lora Weiss

69 78

86

87

89

97

105

Using Web-Based Knowledge Extraction Techniques to Support Cultural Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Paul R. Smart, Winston R. Sieck, and Nigel R. Shadbolt

113

How Corruption Blunts Counternarcotic Policies in Afghanistan: A Multiagent Investigation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Armando Geller, Seyed M. Mussavi Rizi, and Maciej M. L  atek

121

Discovering Collaborative Cyber Attack Patterns Using Social Network Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Haitao Du and Shanchieh Jay Yang

129

Agents That Negotiate Proficiently with People (Keynote) . . . . . . . . . . . . Sarit Kraus Cognitive Modeling for Agent-Based Simulation of Child Maltreatment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Xiaolin Hu and Richard Puddy

137

138

Table of Contents

XI

Crowdsourcing Quality Control of Online Information: A Quality-Based Cascade Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Wai-Tat Fu and Vera Liao

147

Consumer Search, Rationing Rules, and the Consequence for Competition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Christopher S. Ruebeck

155

Pattern Analysis in Social Networks with Dynamic Connections . . . . . . . Yu Wu and Yu Zhang

163

Formation of Common Investment Networks by Project Establishment between Agents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jes´ us Emeterio Navarro-Barrientos

172

Constructing Social Networks from Unstructured Group Dialog in Virtual Worlds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Fahad Shah and Gita Sukthankar

180

Effects of Opposition on the Diffusion of Complex Contagions in Social Networks: An Empirical Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Chris J. Kuhlman, V.S. Anil Kumar, Madhav V. Marathe, S.S. Ravi, and Daniel J. Rosenkrantz Promoting Coordination for Disaster Relief – From Crowdsourcing to Coordination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Huiji Gao, Xufei Wang, Geoffrey Barbier, and Huan Liu Trust Maximization in Social Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Justin Zhan and Xing Fang Towards a Computational Analysis of Status and Leadership Styles on FDA Panels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . David A. Broniatowski and Christopher L. Magee

188

197

205

212

Tracking Communities in Dynamic Social Networks . . . . . . . . . . . . . . . . . . Kevin S. Xu, Mark Kliger, and Alfred O. Hero III

219

Bayesian Networks for Social Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Paul Whitney, Amanda White, Stephen Walsh, Angela Dalton, and Alan Brothers

227

A Psychological Model for Aggregating Judgments of Magnitude . . . . . . . Edgar C. Merkle and Mark Steyvers

236

XII

Table of Contents

Speeding Up Network Layout and Centrality Measures for Social Computing Goals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Puneet Sharma, Udayan Khurana, Ben Shneiderman, Max Scharrenbroich, and John Locke

244

An Agent-Based Simulation for Investigating the Impact of Stereotypes on Task-Oriented Group Formation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Mahsa Maghami and Gita Sukthankar

252

An Approach for Dynamic Optimization of Prevention Program Implementation in Stochastic Environments . . . . . . . . . . . . . . . . . . . . . . . . . Yuncheol Kang and Vittal Prabhu

260

Ranking Information in Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Tina Eliassi-Rad and Keith Henderson

268

Information Provenance in Social Media . . . . . . . . . . . . . . . . . . . . . . . . . . . . Geoffrey Barbier and Huan Liu

276

A Cultural Belief Network Simulator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Winston R. Sieck, Benjamin G. Simpkins, and Louise J. Rasmussen

284

Following the Social Media: Aspect Evolution of Online Discussion . . . . . Xuning Tang and Christopher C. Yang

292

Representing Trust in Cognitive Social Simulations . . . . . . . . . . . . . . . . . . . Shawn S. Pollock, Jonathan K. Alt, and Christian J. Darken

301

Temporal Visualization of Social Network Dynamics: Prototypes for Nation of Neighbors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jae-wook Ahn, Meirav Taieb-Maimon, Awalin Sopan, Catherine Plaisant, and Ben Shneiderman

309

Toward Predicting Popularity of Social Marketing Messages . . . . . . . . . . . Bei Yu, Miao Chen, and Linchi Kwok

317

Simulating the Effect of Peacekeeping Operations 2010–2035 . . . . . . . . . . H˚ avard Hegre, Lisa Hultman, and H˚ avard Mokleiv Nyg˚ ard

325

Rebellion on Sugarscape: Case Studies for Greed and Grievance Theory of Civil Conflicts Using Agent-Based Models . . . . . . . . . . . . . . . . . . . . . . . . Rong Pan

333

A Culture-Sensitive Agent in Kirman’s Ant Model . . . . . . . . . . . . . . . . . . . Shu-Heng Chen, Wen-Ching Liou, and Ting-Yu Chen

341

Modeling Link Formation Behaviors in Dynamic Social Networks . . . . . . Viet-An Nguyen, Cane Wing-Ki Leung, and Ee-Peng Lim

349

Table of Contents

Naming on a Directed Graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Giorgio Gosti and William H. Batchelder Synthesis and Refinement of Detailed Subnetworks in a Social Contact Network for Epidemic Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Huadong Xia, Jiangzhuo Chen, Madhav Marathe, and Henning S. Mortveit

XIII

358

366

Technosocial Predictive Analytics for Illicit Nuclear Trafficking . . . . . . . . Antonio Sanfilippo, Scott Butner, Andrew Cowell, Angela Dalton, Jereme Haack, Sean Kreyling, Rick Riensche, Amanda White, and Paul Whitney

374

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

383

Using Models to Inform Policy: Insights from Modeling the Complexities of Global Polio Eradication Kimberly M. Thompson Kid Risk, Inc., Boston, MA, USA [email protected]

Abstract. Drawing on over 20 years of experience modeling risks in complex systems, this talk will challenge SBP participants to develop models that provide timely and useful answers to critical policy questions when decision makers need them. The talk will include reflections on the opportunities and challenges associated with developing integrated models for complex problems and communicating their results effectively. Dr. Thompson will focus the talk largely on collaborative modeling related to global polio eradication and the application of system dynamics tools. After successful global eradication of wild polioviruses, live polioviruses will still present risks that could potentially lead to paralytic polio cases. This talk will present the insights of efforts to use integrated dynamic, probabilistic risk, decision, and economic models to address critical policy questions related to managing global polio risks. Using a dynamic disease transmission model combined with probabilistic model inputs that characterize uncertainty for a stratified world to account for variability, we find that global health leaders will face some difficult choices, but that they can take actions that will manage the risks effectively. The talk will emphasize the need for true collaboration between modelers and subject matter experts, and the importance of working with decision makers as partners to ensure the development of useful models that actually get used.

J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, p. 1, 2011. © Springer-Verlag Berlin Heidelberg 2011

Toward Culturally Informed Option Awareness for Influence Operations with S-CAT Ken Murray1, John Lowrance1, Ken Sharpe1, Doug Williams2, Keith Grembam2, Kim Holloman3, Clarke Speed4, and Robert Tynes5 1 SRI International, Menlo Park, CA, 94025 USA {kenneth.murray,john.lowrance,kenneth.sharpe}@sri.com 2 SET Corporation, Greenwood Village, CO, 80111 USA {dwilliams,kgrembam}@setcorp.com 3 SAIC, Sterling, VA, 20164 USA [email protected] 4 University of Washington, Seattle, WA 98195, USA [email protected] 5 SUNY at Albany, Albany New York 12222, USA [email protected]

Abstract. The Socio-Cultural Analysis Tool (S-CAT) is being developed to help decision makers better understand the plausible effects of actions taken in situations where the impact of culture is both significant and subtle. We describe S-CAT in the context of a hypothetical influence operation that serves as an illustrative use case. One of the many challenges in developing S-CAT involves providing transparency into the model. S-CAT does this by providing explanations of the analysis it provides. This paper describes how S-CAT can improve option-awareness during influence operations and discusses the explanation capabilities used by S-CAT to support transparency into the model. Keywords: Modeling Culture, Effects-Based Modeling, Forecasting Effects.

1 Introduction The United States and other countries are engaged in Influence Operations (IOs) in very diverse parts of the world. Successful outcomes for such operations require understanding the societies and cultures where they are conducted [1]. Misunderstandings due to cultural differences can yield unforeseen and potentially adverse consequences, delaying progress in the best cases and producing instability and tragedy in the worst cases [2, 3]. To improve effectiveness, IOs must be conducted with improved understanding of the human socio-cultural behavior (HSCB) of the host region in order to pursue plans that avoid otherwise unforeseen adverse consequences and exploit favorable opportunities. Option awareness is improved when selection among candidate plans is informed by an understanding of how they plausibly impact the attitudes, activities and circumstances of the host region. However, not every IO planning analyst can become a social scientist specializing in the host region. J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, pp. 2–9, 2011. © Springer-Verlag Berlin Heidelberg 2011

Toward Culturally Informed Option Awareness for Influence Operations with S-CAT

3

To address this challenge, technology can gather, organize, and represent HSCB knowledge from diverse sources, including social scientists and regional experts, to capture culturally informed action-effect relationships that forecast the plausible consequences of candidate plans [3, 4]. We are developing the Socio-Cultural Analysis Tool (S-CAT) to help analysts and decision makers better understand the plausible effects of actions taken in situations where the impact of culture is both significant and subtle. Although its application is potentially much broader, S-CAT is specifically being developed to provide greater option awareness to planning analysts and decision makers in conducting IOs.

2 Example Scenario: Reducing Piracy Off the West-African Coast The use of S-CAT is illustrated with the hypothetical IO scenario1 in Fig. 1. Following the international crackdown on piracy based in Sudan, Atlantic shipping operations report increasing piracy off the coast of Sierra Leone. Most raiding vessels seem to come from the Bonthe District. The IO coalition must decide how best to curtail this emerging piracy on the Bonthe coast. Four candidate plans are: (1) Provide monetary aid to the Bonthe District Government to discourage piracy in the district. (2) Provide monetary aid to the local tribal leader, Paramount Chief Sesay, to discourage piracy in the jurisdiction of his chiefdom. (3) Provide equipment and supplies (e.g., fish-finding sonar, rice irrigation systems and fertilizer) to the fishing and rice-farming enterprises run by Pa Yangi-Yangi, a local entrepreneur and informal leader. (4) Provide security to Pa Yangi-Yangi, because he appears to be someone the coalition can work with to reduce piracy. Fig. 1. Hypothetical scenario for reducing West-African piracy

IOs often require selecting leaders or organizations from the host region and trying to leverage their local influence to achieve a goal. In the example, candidates include the district government, the local tribal leader, and a local entrepreneur and informal leader. Candidate plans are prepared for each. S-CAT contributes to option awareness during plan selection by: • • •

Forecasting the plausible effects of each alternative plan Comparing the forecasted effects of alternative plans with stated desired effects Revealing the cultural factors and lines of reasoning supporting forecasted effects

The following sections describe how S-CAT uses several types of knowledge to assess the candidate plans in Fig. 1 and how S-CAT is being developed to provide transparency into its model by explaining the results of its analysis. 1

The people and events described are not real. The candidate plans are for purposes of illustration only and do not represent the real or intended actions of any government.

4

K. Murray et al.

3 Improving Option Awareness by Assessing Alternative Plans S-CAT supports two distinct users: a modeling analyst uses S-CAT to construct culturally informed models used to forecast the plausible effects of plans and a planning analyst uses S-CAT to construct and assess candidate plans. 3.1 Modeling in S-CAT The modeling analyst requires access to cultural expertise but is otherwise lightly trained: S-CAT is being developed explicitly to enable use by IO analysts so that extensive training in knowledge representation and other esoteric modeling specialties is not required. The modeling analyst constructs two types of model components: the site model and the effects rules. Sites. The site model captures elements of the area of interest that are reasoned about; this typically includes: • • •

Agents, including individual people, groups (e.g., tribes, populations) and organizations (e.g., governments, NGOs) Locations, including geopolitical regions (e.g., nations, districts, villages) and geographical regions (e.g., islands, rivers, plains) Activities, including economic enterprises (e.g., fishing, farming)

There is no restriction on the echelon or level of detail captured in the site model: international regions and nation states can coexist with individuals and villages. Each element is specified as instantiating one or more types (e.g., Leader, Village) selected from an existing, extensible hierarchy of element types. The site model includes relations over the elements (e.g., relation-type resident-of captures that a person resides in a region). The site also includes profiles that describe the cultural properties of elements. For this example, S-CAT uses two culture profiles, Affluence and Speed ([4] describes the capture and use of cultural profiles and their variables in S-CAT). Fig. 2 presents a portion of the site model developed for the example.

Fig. 2. Part of the site model for reducing West-African piracy

Toward Culturally Informed Option Awareness for Influence Operations with S-CAT

5

Effects rules. S-CAT includes two kinds of effects rules: direct and indirect. Directeffects rules capture effects that plausibly occur directly from the execution of some type of action on some type of site element. Indirect-effects rules capture the effects that plausibly occur as a consequence of some other effect; they reveal how an impact on one element of the site model can propagate to other related elements. The antecedent of a direct-effects rule is simply the action type, the element type, and optional cultural constraints on the site elements to which the rule applies. The antecedent of an indirect-effects rule includes typed “variable nodes” with relations and constraints on the values of effects and cultural variables. The consequences of all effects rules are plausible values for effects variables. This example uses variables capturing the political, military, economic, social, infrastructure, and information (PMESII) effects ([4] describes the use of PMESII effects variables in S-CAT, and [5] describes the structured-argumentation techniques supporting their capture). Fig. 3 presents a direct-effects rule and shows how the modeling analyst specifies a consequence PMESII effect. Fig. 4 presents an indirect-effects rule and its rationale. Action Type: Provide-Industry-Equipment Site-Element Type: Small-Enterprise

Affluence Constraint: Speed Constraint: Direct PMESII Effects:

Fig. 3. Direct-Effects Rule: Provide Industry Equipment to a Small Enterprise

Rationale: When an economic enterprise produces a surplus and its performers embrace allocative leadership, their leader controls that surplus. Fig. 4. Indirect-Effects Rule: Allocative Leader Controls Enterprise Surplus

6

K. Murray et al.

3.2 Assessing Plans in S-CAT After the modeling analyst has created the site model and a set of site-appropriate (culturally informed) effects rules, the planning analyst uses the model to define and assess candidate plans. This involves: creating a goal to capture the desired effects, creating plans to achieve the goal, creating forecasts of plausible effects for each plan, and then evaluating the plans’ forecasted effects against the goal’s target effects. Goals. A goal is defined by selecting one or more elements from the site model, and for each creating a target PMESII profile that identifies the desired PMESII effects for that site element. Each target effect can be annotated with a weight to capture priorities. In the example, the two site elements selected are Kita Piracy and Bonthe District Population. The primary (and fully weighted) target effects include reducing the activity of piracy and reducing its profitability. The secondary target effects include reducing support for piracy among informal leaders and increasing the social stability of the population. Fig. 5 presents the goal defined for the example.

Fig. 5. A counter-piracy goal

Plans. A plan contains one or more actions. Each action is defined as the application of an action type, selected from an existing, extensible taxonomy of diplomatic, information, military, and economic (DIME) action types, to an element selected from the site model. In the example, there are four distinct plans created; the third plan from Fig. 1 is presented in Fig. 6. This plan aspires to leverage the influence of a local entrepreneur, Pa Yangi-Yangi. It includes providing industrial equipment (e.g., fishfinding sonar) to Kita Fishing and providing both industrial equipment (e.g., irrigation equipment, rice-milling machines) and industrial supplies (e.g., fertilizer) to Kita Rice Farming; both operations are controlled by Pa Yangi-Yangi.

Fig. 6. A counter-piracy plan: Invest in Kita Food Production

Toward Culturally Informed Option Awareness for Influence Operations with S-CAT

7

Assessing Alternative Plans. For each candidate plan, the planning analyst uses SCAT to create a forecast of plausible effects from executing the plan. This involves first running all the applicable direct-effects rules and then, iteratively, running all the applicable indirect-effects rules. Each forecast creates a distinct inferential context for the effects it establishes for the site elements and enables S-CAT to maintain multiple, independent forecasts (e.g., one for each candidate plan), where the effects in one forecast do not impact the effects in another forecast. The applicable direct-effects initiate the forecast. A direct-effects rule is applicable for a plan if the rule’s action-type matches or generalizes the action-type specified in one of the plan’s actions, and the site element selected for that plan action instantiates the site-element type of the rule, and that site element satisfies the cultural constraints, if any, defined for the rule. In the example, the direct-effects rule in Fig. 4 applies to the first two actions of the plan in Fig. 6 (the Affluence-profile constraint in Fig. 4 requires the enterprise to use primitive tools and is satisfied by both Kita Fishing and Kita Rice Farming). The applicable indirect-effect rules extend the forecast by iteratively propagating effects established by the forecast (e.g., by previously applied effects rules) to other related site elements. An indirect-effects rule is applicable when each of its variables bind to some site element such that the types for the variable are instantiated by the site element, the relations defined for the variables are satisfied by the site-element bindings, and each site element satisfies any cultural profile constraints defined for the variable as well as any effects constraints defined for the variable. Note that the effects constraints are tested using the forecast, whereas the other conditions are tested using the site model. After forecasts are generated for the alternative plans, they are evaluated against the goal. S-CAT displays a table showing how well the effects of each plan’s forecast satisfy the target effects of the goal (see Fig. 7). According to the model, investing in the local Paramount Chief or the district government has no impact on piracy. Among these four plans, only investing in the local food-production enterprises leads to a reduction in piracy. Furthermore, a quite surprising outcome is that providing Pa Yangi-Yangi protection might actually reduce social stability (because of the unforeseen consequence of putting him at greater risk from rival, militant leaders).

Fig. 7. Comparing forecasted effects among alternative candidate plans

8

K. Murray et al.

4 Improving Option Awareness by Explaining Forecasted Effects S-CAT is not intended to simply identify the best among a set of candidate plans. Rather, it forecasts specific effects and explains the forecasted effects in order to reveal the cultural factors and lines of reasoning that support each forecasted effect. This provides essential transparency into the HSCB model and distinguishes S-CAT from other HSCB tools using traditional agent-based simulation capabilities [6]. Identifying the cultural factors relevant to forecasted effects increases the understanding of the IO planning analysts and decision makers about how candidate plans impact and are impacted by the host region culture, thereby improving their option awareness. By revealing precisely those cultural factors in the model most relevant to a candidate plan as that plan is being evaluated for selection, S-CAT provides just in time cultural training to the IO planning analysts and decision makers. Identifying the lines of reasoning used in a forecast has two primary benefits. It increases the understanding of the IO planning analysts and decision makers about why the forecasted effects may plausibly occur, thereby improving their option awareness. Furthermore, it increases their understanding of what is and what is not captured in the model and considered by S-CAT when computing forecasts, thereby improving their awareness of the model’s scope and limitations. Fig. 8 presents samples of the cultural factors and lines of reasoning revealed by SCAT for the reduction in Kita Piracy forecasted for the plan of Fig. 6.

Fig. 8. Explaining the forecasted reduction in Kita Piracy

Toward Culturally Informed Option Awareness for Influence Operations with S-CAT

9

5 Summary and Discussion S-CAT is a decision-support tool being developed to improve option awareness by providing culturally informed assessments of candidate plans. Each assessment forecasts plausible effects of a plan. Explanation capabilities reveal both the cultural factors and the lines of reasoning supporting each forecasted effect, providing transparency into the model and the automated reasoning used to assess the plans. Revealing the cultural factors enables users to better understand how the local culture impacts their candidate plans. Revealing the lines of reasoning enables users to better understand how the forecasted effects of their candidate plans are plausible. By identifying and explaining both the plausible effects and the underlying cultural factors relevant to candidate plans, S-CAT improves the option awareness of its users. S-CAT does not obviate the need for cultural expertise about the region of interest. Rather, it allows lightly trained modeling analysts to capture cultural expertise provided from available sources (e.g., social scientists, regional experts) in culturally informed models. Once created, the models can be subsequently exploited anywhere and at anytime in support of IO planning. Thus, capturing cultural expertise within SCAT enables the access to available sources of cultural expertise (during modeling) to be asynchronous and offsite from the application of that expertise (during planning). Acknowledgments. This work has benefited from contributions by Ian Harrison, Janet Murdoch, Eric Yeh, and David Anhalt, and it has been supported in part through a contract with the Combating Terrorism Technical Support Office with funding from the U.S. Department of Defense, Office of the Director for Defense Research and Engineering (DDR&E), Human Social Culture and Behavior Modeling Program.

References 1. Ng, K.Y., Ramaya, R., Teo, T.M.S., Wong, S.F.: Cultural Intelligence: Its Potential for Military Leadership Development. In: Proc. of IMTA 2005, Singapore, pp. 154–167 (2005) 2. Hearing: Army Transformation Hearing before the Committee on Armed Services, United States House of Representatives, Second Session, Washington, D.C., July 15-24 (2004) 3. Sharpe, K., Gremban, K., Holloman, K.: Evaluating the Impact of Culture on Planning and Executing Multinational Joint Force Stability, Security, Transition and Reconstruction Operations. In: Proceedings of KSCO 2007, Waltham, MA, pp. 76–81 (2007) 4. Murray, K., Lowrance, J., Sharpe, K., Williams, D., Grembam, K., Holloman, K., Speed, C.: Capturing Culture and Effects Variables Using Structured Argumentation. In: Schmorrow, D., Nicholson, D. (eds.) Advances in Cross-Cultural Decision Making, pp. 363–373. CRC Press, Boca Raton (2010) 5. Lowrance, J., Harrison, I., Rodriguez, A., Yeh, E., Boyce, T., Murdock, J., Murray, K.: Template-Based Structured Argumentation. In: Okada, A., Buckingham Shum, S.J., Sherborne, T. (eds.) Knowledge Cartography: Software Tools and Mapping Techniques, pp. 307–333. Springer, London (2008) 6. Russell, A., Clark, M., La Valley, R., Hardy, W., Zartman, I.W.: Evaluating Human, Social, Cultural, and Behavior (HSCB) Models for Operational Use. In: Schmorrow, D., Nicholson, D. (eds.) Advances in Cross-Cultural Decision Making, pp. 374–384. CRC Press, Boca Raton (2010)

Optimization-Based Influencing of Village Social Networks in a Counterinsurgency Benjamin W.K. Hung1, Stephan E. Kolitz2, and Asuman Ozdaglar3 1

United States Military Academy, Department of Mathematical Sciences, West Point, New York 10996 [email protected] 2 Charles Stark Draper Laboratory, Cambridge, Massachusetts 02139 [email protected] 3 Massachusetts Institute of Technology, Laboratory for Information and Decision Systems, Cambridge, Massachusetts 02139 [email protected]

Abstract. This paper considers the nonlethal targeting assignment problem in the counterinsurgency in Afghanistan, the problem of deciding on the people whom US forces should engage through outreach, negotiations, meetings, and other interactions in order to ultimately win the support of the population in their area of operations. We propose two models: 1) the Afghan COIN social influence model, to represent how attitudes of local leaders are affected by repeated interactions with other local leaders, insurgents, and counterinsurgents, and 2) the nonlethal targeting model, a nonlinear programming (NLP) optimization formulation that identifies a strategy for assigning k US agents to produce the greatest arithmetic mean of the expected long-term attitude of the population. We demonstrate in an experiment the merits of the optimization model in nonlethal targeting, which performs significantly better than both doctrine-based and random methods of assignment in a large network. Keywords: social counterinsurgency.

network,

agent

modeling,

network

optimization,

1 Problem Statement Insurgent movements rely on the broad support of the local population in seeking to overthrow weak governments. The US military finds itself conducting counterinsurgency (COIN) operations on the side of these weak host nation governments. Thus, a counterinsurgency ultimately succeeds by gaining the support of the population, not by killing or capturing insurgents alone [1]. A key component of these counterinsurgency operations is deciding which among a broad class of non-lethal influencing activities ought to be undertaken to win the support of the local population. These activities range from engaging local leaders in constructive dialogue to addressing root causes of poor governance by providing aid and reconstruction assistance for the population. In the counterinsurgency operations in both Afghanistan and Iraq, units of battalions and companies are assigned vast amounts of territory and often charged J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, pp. 10–17, 2011. © Springer-Verlag Berlin Heidelberg 2011

Optimization-Based Influencing of Village Social Networks in a Counterinsurgency

11

with winning the support of tens of thousands of people. These units are resourceconstrained in terms of personnel, money, equipment, and time. These constraints force units to be very selective in targeting the number, type, frequency, and objective of the operations they conduct. Furthermore, because success or failure is contingent on human behavior, it is extremely difficult to predict how a given action will affect a group of people. The current methods of prioritizing and determining target value are qualitative and are based upon the commanders’ and staffs’ intuition. Our objective is to develop a quantitative approach to address how units can determine the best non-lethal targeting assignment strategies and how the sentiments of the population might change as a result of them.

2 Technical Approach The technical approach was developed to support a variety of uses, including: 1) providing decision support to commanders and their staffs when they make decisions on whom to target non-lethally in order to maximize popular support while operating within the unit’s doctrinal and resource constraints, and 2) developing effective COIN Tactics, Techniques and Procedures (TTPs) for non-lethal engagement. We employ the following technical methods: social network analysis, agent modeling, and network optimization. Social network analysis is an interdisciplinary field that models interrelated individuals or groups of people as nodes in a network [2]. For this application, we model the actors (“agents”) in an Afghan counterinsurgency (e.g., local leaders, Taliban and US forces) as nodes and the relationships among them as links in a social network. Agent modeling provides analytic insight into the emergent behavior of a collection of agents (individuals or groups of individuals) based on a set of rules that characterize pair-wise interactions. In particular, each agent in the network has a scalar-valued attitude of favorability towards US forces, and the rules model probabilities of influencing each of its neighbors’ attitudes. Network optimization here refers to formulating and solving the problem of assignment of activities by US forces in order to maximize the positive attitude of local leaders. Our work is related to a large opinion dynamics literature that presents explanatory models of how individual opinions may change over time. See Abelson [3], DeGroot [4], Friedkin and Johnsen [5], Deffuant [6] [7], and Hegselmann and Krause [8]. This work is most closely related to and builds upon the work of Acemoglu, et al. which characterized beliefs in a social network in the midst of influential agents [9], and “stubborn” agents (i.e., agents which do not change belief) [10]. This work is also related to key person problem (KPP) literature, which uses quantitative methods to identify the k-best agents of diffusion of binary behaviors in a network. For example, see Borgatti [11] and Kempe, Kleinburg, and Tardos [12]. From previous work, we make three main contributions. First we apply existing opinion dynamics research to the context of the counterinsurgency in Afghanistan by introducing a methodology for determining approximations of interpersonal influences over Afghan social interaction networks based upon social science and anthropological research. We also provide a method to determine a quantitative value of nonlethal US targeting assignments and their predicted effect over a continuous scale on the local population in the counterinsurgency. Lastly, we present an optimization-based method of nonlethal target assignment selection.

12

B.W.K. Hung, S.E. Kolitz, and A. Ozdaglar

The starting point for our Afghan COIN social influence model is a flexible and extensible Markovian model developed by Acemoglu, et al. [9]. The model represents the agents as nodes in the network graph. Values at each node represent the attitudes of the respective agent and the edges of the graph represent connections among the agents over which influence occurs. The level of influence an agent (node) exerts on other agents (nodes) is modeled by the degree of forcefulness of each, as described below. We model an agent’s attitude towards the counterinsurgents as a continuous random variable that takes on a scalar value at each interaction occurrence (over all the agents). Specifically, let ( ) and ( ) represent the value of node i and j, respectively, at interaction k, where both ( ) and ( ) are between -0.5 and +0.5. A negative (or positive) value means low (or high) favorability towards the counterinsurgents, and zero means neutral. This spectrum of attitudes is depicted in Figure 1. Extreme points along the scale denote a greater strength of attitude [7]. In our model, the ideologically motivated agents, the US counterinsurgents and the Taliban insurgents, are “immutable” agents because they possess immutable attitudes which remain at the extreme points and do not change over time.

Fig. 1. The numeric and gray scale of the population attitudes

Figure 2 illustrates a simple example network of three villages as well as a US forceful agent and Taliban forceful agent. The shade of a node represents the attitude value using the scale above.

Fig. 2. Illustrative network model

Optimization-Based Influencing of Village Social Networks in a Counterinsurgency

13

The state of the system is the set of values (attitudes) for the nodes in the network. Changes in the state of the system are modeled using a Markov chain. Transitions occur when there is an interaction between two nodes. Interactions occur when two nodes are chosen sequentially (to model who approaches whom) out of an agent selection probability distribution across nodes. This meeting probability distribution could be uniform, (i.e., each agent is equally likely to be selected), or non-uniform, (i.e., certain agents are more likely to instigate interactions). The first node selected is denoted as i and the second node is denoted as j. The values for nodes i and j can change in response to the interaction between i and j. The transition from an interaction depends upon the forcefulness (described below) of the agents. In the interaction between two agents with values ( ) and ( ), the possible event outcomes are: 1. With probability , agents i and j reach a pair-wise consensus of attitudes: (

1)

(

( )

1)

( )

.

(1)

2. With probability , agent j exerts influence on agent i, where the parameter measures the tendency for agent i to resist change to his/her attitude in response to the interaction with agent j: ( ( 3. With probability attitude:

· ( ) ( )

1) 1)

1

·

( )

.

(2)

, agents i and j do not agree and each retains their original ( (

1) 1)

( ) . ( )

(3)

1. Note that These interaction-type probabilities are parameters in the model and can depend on the relative forcefulness of the two agents. Forcefulness can take on one of four values1: regular (lowest level of forcefulness), forceful0 (third highest level of forcefulness), forceful1 (second highest level of forcefulness), and forceful2 (highest level of forcefulness). In an exchange between two agents, the agent most likely to change in value is the less forceful of the two and the amount of change depends on the parameter , which can be thought of as a measure of consistency of attitude. The transitions can be symmetric, i.e., the order in which the agents interact (who approaches whom) does not matter, or can be different dependent upon the order of the interaction. The most forceful agents (forceful2) always exert an influence on less forceful agents, although the strength of the influence is parameterized. At the other extreme, regular agents are always influenced by forceful agents. For cases in between, multiple outcomes are possible. Varying the parameters and probabilities provide significant modeling flexibility. See [13] for more details. 1

Clearly, the model could have any number of forceful types, including 1 (as in the Acemoglu et al. paper [9]), or 3, as described here. We chose 3 based on a social science SME’s initial guidance.

14

B.W.K. Hung, S.E. Kolitz, and A. Ozdaglar

3 Optimization Formulation While the pair-wise interactions between two agents in the social influence model are simple, the resultant behavior of the system with many agents in a large network is quite complex. However, we devised a method to compute the expected long-term attitudes for each agent analytically. See Section 3.3.4 in [13]. We found that: ·

·

0

(4)

where is the vector of expected long-term attitudes of agents (mutable agents) is the vector of expected long-term attitudes of agents at the interaction as ∞, (immutable agents) at the interaction as ∞ , and the sub-matrices and are parameterized by data of meeting and interaction-type probabilities. The expected long-term attitudes of the mutable agents (immutable agent attitudes do not change) is (

·

).

(5)

This result is based upon a particular topology and parameterization. A natural question that follows is what topology produces the attitudes most favorable to the counterinsurgents? More specifically, how should the US agents form connections to other agents in the network that maximizes the favorable attitudes of the population? In order to answer this question we formulate the nonlethal targeting problem as a nonlinear program (NLP). Drawing from the general methodology of the classical static weapon-target assignment (WTA) problem, we seek to find the assignment of a fixed number of US agents to a fixed number of local leaders in a social network that maximizes the expected long-term attitudes of the population in favor of the US forces. While the complete derivation for this formulation is in Section 3.5 in [13], we explain the key components of the optimization formulation below. Decision variable. Given a fixed number of US agents with (possibly) different interaction-type probabilities, the decision one makes is which US agents are assigned to which non-US agents (local leaders or Taliban) in the network to connect with. The decision variable is therefore a binary matrix, where each entry is: 1, if US Agent connects to agent 0, otherwise

.

(6)

Each US agent can form a link with either 1) the mutable local leaders, or 2) the immutable Taliban leaders. A link formed between a US agent to any mutable agent (and the subsequent propagation of influence from the US agent to that agent) can be representative of any activity or communication in which the targeted local leader is frequently reinforced with pro-counterinsurgent attitudes. See Section 2.5.1 in [13]. A link formed between a US agent and any immutable Taliban agent alters the meeting probabilities with which the Taliban agent negatively influences others. Objective function. We maximize the weighted average of the expected long-term attitudes for all mutable agents in the network, where the unit commander can subjectively determine a weight (value) for each local leader’s attitude based on a variety of factors including 1) the tactical importance of the village where the local leaders are from, and 2) political factors that may demand that one portion of the population be aligned earlier or in lieu of others.

Optimization-Based Influencing of Village Social Networks in a Counterinsurgency

15

Constraints. The constraints are provided by a reformulation of (4) above in terms of the decision variables, and a limit on the number of allowable connections for the -th US agent. This limitation may be based off the leader’s assessment of his or her ability (with limited resources) to reach a certain number of local leaders. While this optimization formulation follows from the analytic expression for the expected long-term attitudes of the population (5), some of its properties make it difficult to solve exactly. Specifically, the formulation is both nonlinear as well as nonconvex, which requires heuristic methods to solve and often arrives only at local optima. Additionally, there are ( ) variables and ( ) constraints, where is the total number of agents in the network, which means the problem is very large. In Section 3.5.6 of [13], we simplify the formulation which reduced the number of variables and constraints to ( ).

4 Results The model has been instantiated in a Monte Carlo Simulation. The starting values { (0), for all i} are generated, based on a priori information or values chosen for analysis purposes. Then a pair of agents is selected using the agent selection probability distribution described previously, and the attitudes are updated according to the transition probabilities described above. The full process can be repeated many times with different starting pairs of nodes and realized transitions. Statistics on the node values can be calculated from results from multiple runs of the model. Example output is illustrated in Section 4.5.2 in [13].

Fig. 3. Experiment network (with village labels)

The input data for the analysis presented here is based on social science aspects of Pashtun society and publicly available aggregate data on Pashtun districts (see [13] for details). A network of 73 local leaders is analyzed, with local leaders being those individuals within the population who by virtue of their authority, power, or position have influence over the attitudes of a group of people. Taliban and US agents are exogenous and exert influence on the network. The 76-node network shown in Figure 3 includes 3 Taliban agents, each with the ability to have connections to 4 local leaders. The network topology is similar that of the “islands” economic model of network formation developed by Jackson [2].

16

B.W.K. Hung, S.E. Kolitz, and A. Ozdaglar

Results from optimization-based selection of the best agents to influence by US forces are compared both analytically as well as in simulation with two alternative methods. The first alternative method is to select connections randomly, choosing only from forceful0 and forceful1 agents. The second method was to select connections based on doctrine found in the US counterinsurgency field manuals including the following principles: • • • •

“Identify leaders who influence the people at the local, regional, and national levels ([1]: 5-9). “Win over “passive or neutral people” ([1]: 5-22) “[Nonlethal targets include] people like community leaders and those insurgents who should be engaged through outreach, negotiation, meetings, and other interaction” ([1]: 5-30). “Start easy… don’t go straight for the main insurgent stronghold …or focus efforts on villages that support the insurgents. Instead, start from secure areas and work gradually outwards. Do this by extending your influence through the locals’ own networks” ([1]: A-5).

Each of the selection methods produced modified network topologies (between US agents and their targeting assignments). Compared to both the random- and doctrinebased selection methods, the optimization-based method produced higher mean expected long-term attitudes both analytically and in simulation in all cases. The full set of experimental results is described in Section 4.5.2 in [13]. The analytical evaluation included a test of the statistical significance of the performance of the optimization-based selection method over the random or doctrine-based selection methods. We tested the null hypothesis: the arithmetic mean of the expected longterm attitudes produced from optimization-based selection and random selection across the cases were drawn from identical continuous distributions with equal medians. We then conducted a pair-wise Wilcox Rank Sum (non-parametric) test on the arithmetic mean of the expected attitudes obtained. The test resulted in a p-value of 0.0041, and thus provided statistically significant evidence that the null hypothesis should be rejected.

Fig. 4. Averaged Mean Expected Long-Term Attitudes

Optimization-Based Influencing of Village Social Networks in a Counterinsurgency

17

Simulations on the resulting network topologies produced plots of the arithmetic mean of the expected long-term attitude at each interaction, k, for k = 1 to 10000 as illustrated in Figure 4. Specifically, we conducted 50 realizations for each selection method within each case. Each realization produced an arithmetic mean of the expected long-term attitude (over all agents in the network) for each interaction, k. We then averaged over those realizations to produce the arithmetic mean (over all realizations) of the arithmetic mean (over all agents) of the expected long-term attitudes for each interaction, k. For simplicity, we will refer to these calculations as the averaged mean expected long-term attitudes. NLTP is the notation for the optimization-based method. We observed that US agents were often assigned to the friendly forceful1 local leaders who had a generally positive (pro-counterinsurgent) effect on attitudes of all village-level leaders connected to them. However, it is also interesting to note that US agents were also occasionally assigned to forceful0 agents in lieu other forceful1 agents (those with higher relative influence). This was likely due to two apparent principles: 1) the need to overcome Taliban influence in certain unfriendly villages, 2) the careful selection of agents to distribute US influence while not being redundant with other US influence efforts.

References 1. Field Manual 3-24, Counterinsurgency. Department of the Army, Washington DC (2006) 2. Jackson, M.: Social and Economic Networks. Princeton University Press, Princeton (2008) 3. Abelson, R.: Mathematical Models of the Distribution of Attitudes Under Controversy. In: Frederiksen, N., Gulliksen, H. (eds.) Contributions to Mathematical Psychology, pp. 141– 160. Holt, Rinehart (1964) 4. DeGroot, M.: Reaching a Consensus. J. Amer. Statistical Assoc. 69(345), 118–121 (1974) 5. Friedkin, N., Johnsen, E.: Social Influence Networks and Opinion Change. Adv. Group Processes 16, 1–29 (1999) 6. Deffuant, G., Neau, D., Amblard, F.: Mixing Beliefs Among Interacting Agents. Adv. Complex Syst. 3, 87–98 (2000) 7. Deffuant, G., Amblard, F., Weisbuch, G., Faure, T.: How can extremism prevail? A study on the relative agreement interaction model. J. Art. Soc. Social Sim. 5(4) (2002) 8. Hegselmann, R., Krause, U.: Opinion Dynamics and Bounded Confidence Models: Analysis and Simulation. J. Art. Soc. Social Sim. 5(3) (2002) 9. Acemoglu, D., Ozdaglar, A., ParandehGheibi, A.: Spread of (Mis)Information in Social Networks. MIT LIDS report 2812, to appear in Games Econ. Behav. (2010) 10. Acemoglu, D., Como, G., Fagnani, F., Ozdaglar, A.: Opinion fluctuations and persistant disagreement in social networks. Submitted to Ann. Appl. Probab. (2010) 11. Borgatti, S.: Identifying sets of key players in a social network. Comp. Math. Org. Theory 12, 21–34 (2006) 12. Kempe, D., Kleinberg, J., Tardos, E.: Maximizing the Spread of Influence through a Social Network. SIGKDD (2003) 13. Hung, B.: Optimization-Based Selection of Influential Agents in a Rural Afghan Social Network. Master’s Thesis, MIT, Operations Research Center, Thesis Supervisor, Asuman Ozdaglar (June 2010)

Identifying Health-Related Topics on Twitter An Exploration of Tobacco-Related Tweets as a Test Topic Kyle W. Prier1, Matthew S. Smith2 , Christophe Giraud-Carrier2, and Carl L. Hanson1 1

Department of Health Science Department of Computer Science Brigham Young University, Provo UT 84602, USA {kyle.prier,smitty,carl hanson}@byu.edu, [email protected] 2

Abstract. Public health-related topics are difficult to identify in large conversational datasets like Twitter. This study examines how to model and discover public health topics and themes in tweets. Tobacco use is chosen as a test case to demonstrate the effectiveness of topic modeling via LDA across a large, representational dataset from the United States, as well as across a smaller subset that was seeded by tobacco-related queries. Topic modeling across the large dataset uncovers several public health-related topics, although tobacco is not detected by this method. However, topic modeling across the tobacco subset provides valuable insight about tobacco use in the United States. The methods used in this paper provide a possible toolset for public health researchers and practitioners to better understand public health problems through large datasets of conversational data. Keywords: Data Mining, Public Health, Social Media, Social Networks, LDA, Tobacco Use, Topic Modeling.

1

Introduction

Over recent years, social network sites (SNS) like Facebook, Myspace, and Twitter have transformed the way individuals interact and communicate with each other across the world. These platforms are in turn creating new avenues for data acquisition and research. Such web-based applications share several common features [3], and while there are slight variations in actual implementations, each service enables users to (1) create a public profile, (2) define a list of other users with whom they share a connection, and (3) view and discover connections between other users within the system. Since SNS allow users to visualize and make public their social networks, this promotes the formation of new connections among users because of the social network platform [3,10]. Not only do SNS enable users to communicate with other users with whom they share explicit social connections, but with a wider audience of users with whom they would not have otherwise shared a social connection. Twitter, in particular, provides a medium whereby users can create and exchange user-generated content with a potentially larger audience than either Facebook or Myspace. J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, pp. 18–25, 2011. c Springer-Verlag Berlin Heidelberg 2011 

Identifying Health-Related Topics on Twitter

19

Twitter is a social network site that allows users to communicate with each other in real-time through short, concise messages (no longer than 140 characters), known as “tweets.” A user’s tweets are available to all of his/her “followers,” i.e., all others who choose to subscribe to that user’s profile. Twitter continues to grow in popularity. As of August 2010, a rough estimate of 54.5 million people used Twitter in the United States, with 63% of those users under the age of 35 years (45% were between the ages of 18 and 34) [17]. In addition, the US accounts for 51% of all Twitter users worldwide. Because Twitter status updates, or tweets, are publicly available and easily accessed through Twitter’s Application Programming Interface (API), Twitter offers a rich environment to observe social interaction and communication among Twitter users [15]. Recent studies have begun to use Twitter data to understand “real,” or offline, health-related behaviors and trends. For example, Chew and Eysenbach analyzed tweets in an effort to determine the types and quality of information exchanged during the H1N1 outbreak [6]. The majority of Twitter posts in their study were news-related (46%), with only 7 of 400 posts containing misinformation. Scanfield et al. explored Twitter status updates related to antibiotics in an effort to discover evidence of misunderstanding and misuse of antibiotics [14]. Their research indicated that Twitter offers a platform for the sharing of health information and advice. In an attempt to evaluate health status, Culotta compared the frequency of influenza-related Twitter messages with influenza statistics from the Centers for Disease Control and Prevention (CDC) [7]. Findings revealed a .78 correlation with the CDC statistics suggesting that monitoring tweets might provide cost effective and quicker health status surveillance. These studies are encouraging. It is our contention that user-generated content on Twitter does indeed provide both public and relevant health-related data, whose identification and analysis adds value to researchers, and may allow them to tailor health interventions more effectively to their audiences. In this paper, we address the problem of how to effectively identify and browse health-related topics within Twitter’s large collection of conversational data. Although recent studies have implemented topic modeling to process Twitter data, these studies have focused on identifying high frequency topics to describe trends among Twitter users. Current topic modeling methods prove difficult to detect lower frequency topics that may be important to investigators. Such methods depend heavily on the frequency distribution of words to generate topic models. Public health-related topics and discussions use less frequent words and are therefore more difficult to identify using traditional topic modeling. Our focus, here, is on the following questions: – How can topic modeling be used to most effectively identify relevant public health topics on Twitter? – Which public health-related topics, specifically tobacco use, are discussed among Twitter users? – What are common tobacco-related themes? We choose to test our topic modeling effectiveness by focusing on tobacco use in the United States. Although we are interested in devising a method to identify

20

K.W. Prier et al.

and better understand public health-related tweets in general, a test topic provides a useful indicator through which we can guide and gauge the effectiveness of our methodology. Tobacco use is a relevant public health topic to use as a test case, as it remains one of the major health concerns in the US. Tobacco use, primarily cigarette use, is one of the leading health indicators of 2010 as determined by the Federal Government [11]. Additionally, tobacco use is considered the most preventable cause of disease and has been attributed to over 14 million deaths in the United States since 1964 [16,1]. Also, there remain approximately 400,000 smokers and former smokers who die each year from smoking-related diseases, while 38,000 non-smokers die each year due to second-hand smoke [1,12]. In the following sections, we discuss our data sampling and analysis of tweets. We used Latent Dirichlet Allocation (LDA) to analyze terms and topics from the entire dataset as well as from a subset of tweets created by querying general, tobacco use-related terms. We highlight interesting topics and connections as well as limitations to both approaches. Finally, we discuss our conclusions regarding our research questions as well as possible areas of future study.

2

Methods and Results

In this section, we introduce our methods of sampling and collecting tweets. Additionally, we discuss our methods to analyze tweets through topic modeling. We used two distinct stages to demonstrate topic modeling effectiveness. First, we model a large dataset of raw tweets in order to uncover health-related issues. Secondly, we create a subset of tweets by querying a raw dataset with tobaccorelated terms. We run topic modeling on this subset as well, and we report our findings. 2.1

Data Sampling and Collection

In order to obtain a representative sample of tweets within the United States, we chose a state from each of the nine Federal Census divisions through a random selection process. For our sample, we gathered tweets from the following states: Georgia, Idaho, Indiana, Kansas, Louisiana, Massachusetts, Mississippi, Oregon, and Pennsylvania. Using the Twitter Search API, recent tweets for each state were gathered in two minute intervals over a 14-day period from October 6, 2010 through October 20, 2010. We used the Twitter Search API’s “geocode” parameter to retrieve the most recent tweets from areas within each of the states, regardless of the content of the tweets. Since the “geocode” parameter requires a circle radius length, latitude and longitude, we gathered tweets from multiple specified circular areas that spanned each state’s boundaries. To prepare the dataset for topic modeling, we remove all non-latin characters from the messages, replace all links within the dataset with the term “link,” and only include users that publish at least two tweets. This process results in a dataset of 2,231,712 messages from 155,508 users. We refer to this dataset as the “comprehensive” dataset.

Identifying Health-Related Topics on Twitter

2.2

21

Comprehensive Dataset Analysis

We analyze the comprehensive dataset by performing LDA on it to produce a topic model. LDA is an unsupervised machine learning generative probabilistic model, which identifies latent topic information in large collections of data including text corpora. LDA represents each document within the corpora as a probability distribution over topics, while each topic is represented as a probability distribution over a number of words [2,9]. LDA enables us to browse words that are frequently found together, or that share a common connection, or topic. LDA helps identify additional terms within topics that may not be directly intuitive, but that are relevant. We configure LDA to generate 250 topic distributions for single words (unigrams) as well as more extensive structural units (n-grams), while removing stopwords. We began at 100 topics and increased the topic number by increments of 50. At 250 topics, the model provided topics that contained unigrams and n-grams that exhibited sufficient cohesion. Additionally, we set LDA to run for 1,000 iterations. We suspected that this topic model would provide less-relevant topics relating to public health and specifically tobacco use, since such topics are generally less frequently discussed. The LDA model of the comprehensive dataset generally demonstrates topics related to various sundry conversational topics, news, and some spam as supported by Pear Analytics [5]. However, the model provides several topics, which contain health-related terms as shown in Table 1. The table includes a percent of the tweets that used any of the n-grams or unigrams within each topic. Although counting tweets that have any of the topic terms may be considered generous, we decided that within the study’s context this measurement provides a relatively simple metric to evaluate the frequency of term usage. Several themes are identifiable from the LDA output: physical activity, obesity, substance abuse, and healthcare. Through these topics we observe that those relating to obesity and weight loss primarily deal with advertising. Additionally, we observe that the terms relating to healthcare refer to current events and some political discourse. Such an analysis across a wide range of conversational data demonstrates the relative high frequency of these health-related topics. However, as we suspected, this analysis fails to detect lower frequency topics, specifically our test topic, tobacco use. In order to “bring out” tobacco use topics, we create a smaller, more focused dataset that we will refer to as the “tobacco subset.” 2.3

Tobacco Subset Analysis

To build our “tobacco subset,” we query the comprehensive dataset with a small set of tobacco-related terms: “smoking,” “tobacco,” “cigarette,” “cigar,” “hookah,” and “hooka.”We chose these terms because they are relatively unequivocal, and they specifically indicate tobacco use. We chose the term “hookah” because of an established trend, particularly among adolescents and college students, to engage in hookah, or waterpipe, usage [8]. The recent emergence of hookah bars provides additional evidence of the popularity of hookah smoking [13].

22

K.W. Prier et al.

Table 1. Comprehensive Dataset Topics Relevant to Public Health. The last column is the percent of tweets that used any of the n-grams or unigrams within each topic. Topic Most Likely Topic Components (n-grams) 44 gps app, calories burned, felt alright, race report, weights workout, christus schumpert, workout named, chrissie wellington, started cycling, schwinn airdyne, core fitness, vff twitter acct, mc gold team, fordironman ran, fetcheveryone ive, logyourrun iphone, elite athletes, lester greene, big improvement, myrtle avenue 45 alzheimers disease, breast augmentation, compression garments, sej nuke panel, weekly newsletter, lab result, medical news, prescription medications, diagnostic imaging, accountable care, elder care, vaser lipo, lasting legacy, restless legs syndrome, joblessness remains, true recession, bariatric surgery, older applicants, internships attract, affordable dental 131 weight loss, diet pills, acai berry, healthy living, fat loss, weight loss diets, belly fat, alternative health, fat burning, pack abs, organic gardening, essential oils, container gardening, hcg diet, walnut creek, fatty acids, anti aging, muscle gain, perez hilton encourages Topic Most Likely Topic Components (unigrams) 18 high, smoke, shit, realwizkhalifa, weed, spitta, currensy, black, bro, roll, yellow, man, hit, wiz, sir, kush, alot, fuck, swag, blunt

% 0.01

0.01

0.04

% 13.63

Querying the comprehensive dataset with these terms, results in a subset of 1,963 tweets. The subset is approximately 0.1% of the comprehensive dataset. After running LDA on this subset, we found that there were insufficient data to determine relevant tobacco-related topics. We thus extended the tobacco subset to include tweets that were collected and preprocessed in the same manner as the comprehensive dataset. The only difference is that we used tweets collected from a 4-week period between October 4, 2010 and November 3, 2010. This resulted in a larger subset of 5,929,462 tweets, of which 4,962 were tobacco-related tweets (approximately 0.3%). Because of the limited size of the subset, we reduce the number of topics returned by LDA to 5, while we retain the output of both unigrams and n-grams. The topics displayed in Table 2, contain several interesting themes relating to how Twitter users discuss tobacco-related topics. Topic 1 contains topics related not only to tobacco use, but also terms that relate to substance abuse including marijuana and crack cocaine. Topic 2 contains terms that relate to addiction recovery: quit smoking, stop smoking, quitting smoking, electronic cigarette, smoking addiction, quit smoking cigarettes, link quit smoking, and link holistic remedies. Topic 3 is less cohesive, and contains terms relating to addiction recovery: quit smoking, stop smoking, secondhand smoke, effective steps. Additionally, it contains words relating to tobacco promotion by clubs or bars: drink specials, free food, ladies night. Topic 4 contains terms related to both promotion by bars or clubs (ladies free, piedmont cir, hookahs great food, smoking room, halloween party, million people) and marijuana use (smoking weed, pot smoking). Topic 5 contains several terms that relate to anti-smoking

Identifying Health-Related Topics on Twitter

23

Table 2. Tobacco Subset Topics. The most likely topic components for each of the five topics generated by LDA. The last column is the percent of tweets that used any of the n-grams within each topic. Topic Most Likely Topic Components (n-grams) 1 smoking weed, smoking gun, smoking crack, stop smoking, cigarette burns, external cell phones, hooka bar, youre smoking, smoke cigars, smoking kush, hand smoke, im taking, smoking barrels, hookah house, hes smoking, ryder cup, dont understand, talking bout, im ready, twenty years people 2 quit smoking, stop smoking, cigar guy, smoking cigarettes, hookah bar, usa protect, quitting smoking, started smoking, electronic cigarette, cigars link, smoking addiction, cigar shop, quit smoking cigarettes, chronical green smoke, link quit smoking naturally, smoking pot, youtube video, link quit smoking, link holistic remedies, chronical protect 3 cigarette smoke, dont smoke, quit smoking, stop smoking, smoking pot, im gonna, hookah tonight, smoking ban, drink specials, free food, ladies night, electronic cigarettes, good times, smoking session, cigarette break, secondhand smoke, everythings real, effective steps, smoking cigs, smoking tonight 4 smoking weed, cont link, ladies free, piedmont cir, start smoking, hate smoking, hookahs great food, cigarette butts, thingswomenshouldstopdoing smoking, lol rt, sunday spot, cigarettes today, fletcher knebel smoking, pot smoking, film stars, external cell, fetishize holding, smoking room, halloween party, million people 5 smoke cigarettes, smoking hot, im smoking, smoking section, stopped smoking, chewing tobacco, smoking kills, chain smoking, smoking area, ban smoking, people die, ring ring hookah ring ring, love lafayette, link rt, damn cigarette, healthiest smoking products, theyre smoking, hate cigarettes, world series, hideout apartment

% 0.16

0.27

0.22

0.25

0.06

and addiction recovery themes: stopped smoking, smoking kills, chain smoking, ban smoking, people die, damn cigarette, hate cigarettes. In this case, topic modeling has helped to understand more fully how users are using tobaccorelated tweets.

3

Discussion and Conclusion

In this study, we directly addressed the problem of how to effectively identify and browse health-related topics on Twitter. We focus on our test topic, tobacco use, throughout the study to explore the realistic application and effectiveness of LDA to learn more about health topics and behavior on Twitter. As expected, we determined that implementing LDA over a large dataset of tweets provides very few health topics. The health topics that LDA does produce during this first stage suggests the popularity of these topics in our dataset. The topic relating to weight loss solutions indicate a high frequency of advertisements in this area.

24

K.W. Prier et al.

The topic relating to healthcare, Obama, and other current health issues indicate a current trend in political discourse related to health. Additionally, the high frequency of marijuana-related terms indicates a potentially significant behavior risk that can be detected through Twitter. While this method did not detect lower frequency topics, it may still provide public health researchers insight into popular health-related trends on Twitter. This suggests that there is potential research needed to test LDA as an effective method to identify health-related trends on Twitter. The first method can provide answers to research questions regarding what topics in general are most discussed among Twitter users. The second method we used to identify tobacco-related topics appeares to be most promising to identify and understand public health topics. Although this method is less automated and requires us to choose terms related to tobacco use, the results indicated this method may be a valuable tool for public health researchers. Based on the results of our topic model, Twitter has been identified as a potentially useful tool to better understand health-related topics, such as tobacco. Specifically, this method of Twitter analysis enables public health researchers to better monitor and survey health status in order to solve community health problems [4]. Our results suggest that chronic health behaviors, like tobacco use, can be identified and measured over shorter periods of time. However, our results do not verify the extent to which short term health events like disease outbreaks can or should be surveyed with this methodology, since tobacco use was used as a test topic. Also, we suspect that the demographics of Twitter users may affect the extent to which topics are discussed on Twitter. We suspect that different public health-related topics may be more or less frequently used on Twitter, and we recommend additional study in this area. Because LDA generates relevant topics relating to tobacco use, we are able to determine themes and also the manner in which tobacco is discussed. In this way, irrelevant conversations can be removed, while tweets related to health status can be isolated. Additionally, by identifying relevant tweets to monitor health status, public health professionals are able to create and implement health interventions more effectively. Researchers can collect almost limitless Twitter data in their areas that will provide practitioners with useful, up-to-date information necessary for understanding relevant public health issues and creating targeted interventions. Finally, the results from the second method suggest that researchers can better understand how Twitter, a popular SNS, is used to promote both positive and negative health behaviors. For example, Topic 4 contains terms that indicate that establishments like bars, clubs, and restaurants use Twitter as a means to promote business as well as tobacco use. In contrast, Topic 2 contains words that relate to addiction recovery by promoting programs that could help individuals quit smoking. The use of LDA in our study demonstrates its potential to extract valuable topics from extremely large datasets of conversational data. While the method proves a valuable outlet to automate the process of removing irrelevant information and to hone in on desired data, it still requires careful human intervention to select query terms for the construction of a relevant subset, and subsequent

Identifying Health-Related Topics on Twitter

25

analysis to determine themes. Research is required to further automate this process. In particular, new methods that can identify infrequent, but highly relevant topics (such as health) among huge datasets will provide value to public health researchers and practitioners, so they can better identify, understand, and help solve health challenges.

References 1. Armour, B.S., Woolery, T., Malarcher, A., Pechacek, T.F., Husten, C.: Annual Smoking-Attributable Mortality, Years of Potential Life Lost, and Productivity Losses. Morbidity and Mortality Weekly Report 54, 625–628 (2005) 2. Blei, D.M., Ng, A.Y., Jordan, M.I.: Latent Dirichlet Allocation. Journal of Machine Learning Research 3, 993–1022 (2003) 3. Boyd, D.M., Ellison, N.B.: Social Network Sites: Definition, History, and Scholarship. Journal of Computer-Mediated Communication 13, 210–230 (2008) 4. Centers for Disease Control and Prevention, http://www.cdc.gov/od/ocphp/nphpsp/essentialphservices.htm 5. Pear Analytics, http://www.pearanalytics.com/blog/2009/twitter-study-revealsinteresting-results-40-percent-pointless-babble/ 6. Chew, C.M., Eysenbach, G.: Pandemics in the Age of Twitter: Content Analysis of “tweets” During the, H1N1 Outbreak. Public Library of Science 5(11), e14118 (2010) (Paper presented 09/17/09 at Medicine 2.0, Naastricht, NL) 7. Culotta, A.: Towards Detecting Influenza Epidemics by Analyzing Twitter Messages. In: Proceedings of the KDD Workshop on Social Media Analytics (2010) 8. Eissenberg, T., Ward, K.D., Smith-Simone, S., Maziak, W.: Waterpipe Tobacco Smoking on a U.S. College Campus: Prevalence and Correlates. Journal of Adolescent Health 42, 526–529 (2008) 9. Griffiths, T.L., Steyvers, M.: Finding Scientific Topics. Proceedings of the National Academy of Sciences 101, 5228–5235 (2004) 10. Haythornthwaite, C.: Social Networks and Internet Connectivity Effects. Information, Communication, & Society 8, 125–147 (2005) 11. Healthy People (2010), http://www.healthypeople.gov/lhi/ 12. Mokdad, A.H., Marks, J.S., Stroup, D.F., Gerberding, J.L.: Actual Causes of Death in the United States. Journal of the American Medical Association 291, 1238–1245 (2004) 13. Primack, B.A., Aronson, J.D., Agarwal, A.A.: An Old Custom, a New Threat to Tobacco Control. American Journal of Public Health 96, 1339 (2006) 14. Scanfield, D., Scanfield, V., Larson, E.: Dissemination of Health Information through Social Networks: Twitter and Antibiotics. American Journal of Infection Control 38, 182–188 (2010) 15. Twitter API documentation, http://dev.twitter.com/doc 16. U.S. Department of Health and Human Services. The Health Consequences of Smoking: A Report for the Surgeon General. Report, USDHHS, Centers for Disease Control and Prevention, National Center for Chronic Disease Prevention and Health Promotion, Office on Smoking and Health (2004) 17. Quantcast, http://www.quantcast.com/twitter.com

Application of a Profile Similarity Methodology for Identifying Terrorist Groups That Use or Pursue CBRN Weapons Ronald L. Breiger1, Gary A. Ackerman2, Victor Asal3, David Melamed1, H. Brinton Milward1, R. Karl Rethemeyer3, and Eric Schoon1 1

2

University of Arizona START Center, University of Maryland 3 University at Albany

Abstract. No single profile fits all CBRN-active groups, and therefore it is important to identify multiple profiles. In the analysis of terrorist organizations, linear and generalized regression modeling provide a set of tools to apply to data that is in the form of cases (named groups) by variables (traits and behaviors of the groups). We turn the conventional regression modeling “inside out” to reveal a network of relations among the cases on the basis of their attribute and behavioral similarity. We show that a network of profile similarity among the cases is built in to standard regression modeling, and that the exploitation of this aspect leads to new insights helpful in the identification of multiple profiles for actors. Our application builds on a study of 108 Islamic jihadist organizations that predicts use or pursuit of CBRN weapons. Keywords: Networks, profile similarity, CBRN use or pursuit, terrorism.

1 Introduction Use or pursuit of chemical, biological, radiological and nuclear weapons (CBRN) on the part of violent non-state actors (VNSA) has been a significant issue for policy makers, academics and the public alike for more than a decade. While concern over the dissemination of chemical and biological weapons predates their use during World War I, Ackerman [1] notes that the widespread contemporary attention to the use and pursuit of weapons capable of inflicting mass casualties was catalyzed in 1995 when the Aum Shinrikyo cult released sarin chemical nerve agent in the Tokyo subway system. Concerns over mass casualty attacks increased dramatically following the jihadist attacks on the World Trade Center on September 11, 2001 [2]. The unprecedented expansion of research and policy initiatives aimed at combating mass casualty attacks, coupled with a growing body of scholarly literature which hypothesizes that radical Islamist organizations—particularly those with a stated jihadist mission—are among the groups most likely to pursue unconventional weapons, has led to an increased focus on predicting the use and pursuit of CBRN among these ideologically motivated organizations, including the use of rigorous multiple regression methods [3,4]. The emphasis on predicting radical Islamist groups’ involvement with CBRN has been J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, pp. 26–33, 2011. © Springer-Verlag Berlin Heidelberg 2011

Application of a Profile Similarity Methodology

27

validated by credible evidence of their interest in these weapons [5]. An overarching question is where analytic attention should be focused [6]. We intend this paper as an illustrative demonstration of the kinds of analysis that are possible when standard regression models are viewed in a new light, one that illuminates the network of profile similarity among the cases that is assumed by the regression models even as it is typically invisible to the analysts. In this paper we illustrate our analytical approach by using it to build on the analysis of Asal and Rethemeyer [3], as their work and the data they utilize stands as an exemplar of rigorous quantitative analysis of CBRN use and pursuit. We recognize nonetheless that a more solid database is necessary in order to advance research that can inform policy. Indeed, a primary objective of our research project is the substantial enhancement of open-source structured databases on CBRN activities by sharply increasing the number of CBRN events coded as well as the number of variables tracked for each event, by developing and applying stronger methods for validating events and assessing the reliability of knowledge of each coded occurrence, and by allowing for longitudinal analysis.

2 Data and Measures Asal and Rethemeyer’s analysis of CBRN use and pursuit encompassed 108 Islamic jihadist groups of which 14 were coded as having pursued or used CBRN weapons.1 As the outcome was coded as binary, logistic regression was employed. Two data sources on CBRN activities were leveraged. The Monterey WMD Terrorism Database, widely regarded as the most comprehensive unclassified data source on the use of CBRN agents by non-state actors, provided data on the use and attempted use of CBRN materials by VNSAs. The Monterey data were combined with data from the National Memorial Institute for the Prevention of Terrorism (MIPT) Terrorism Knowledge database (TKB), and additional data was collected by Asal and Rethemeyer. The data are complete for the years 1998 to 2005. The predictor variables capture features of the organizations as well as of their environments. At the organizational level, membership is a four-category order-ofmagnitude measure of group size ranging from under 100 members to over 10,000. Age is number of years from the group’s inception to 2005, and age-squared is also included. Eigenvector centrality is applied to data (from the “related groups” module of the TKB; see [3] for further discussion) on the alliance connections that terrorist organizations have established among themselves. This measure of centrality recursively weights the number of alliance connections of each organization by the centrality of the groups with which it is connected. Experience is a dummy variable for groups known to have carried out at least four previous attacks (of any type). Control of territory is a dummy variable for groups that maintain control of extensive territory in their “host” country. Turning now to environmental variables, state sponsorship refers to whether (1) or not (0) an organization is coded as being substantially backed by the resources of a 1

Ten organizations considered and pursued CBRN terrorism to some degree but were unsuccessful in actually perpetrating an attack. Four other organizations succeeded at least once in deploying CBRN materials in an attack, none of which killed a large number of people [3].

28

R.L. Breiger et al.

state. The remaining environmental variables pertain to the country in which each organization has its home base. Democracy (POLITY2) is an assessment of regime type on a scale ranging from strongly autocratic (-10) to strongly democratic (+10). Energy per capita, energy consumption relative to total population, taps the level of economic development. To account for degree of economic embeddedness in the global West, trade data between the organization’s home base country and the United States (logged exports to US) were employed. In an effort to measure a country’s link to Western culture, a count of McDonald’s restaurants in each home country (McDonald’s) was obtained. Civil strife is the number of years of civil strife (whether isolated to the home country or internationalized) reported for the study period. Much related information about these variables is given in [3].

3 Profile Similarity We begin with linear regression and then provide some generalization to logistic regression. Consider an n × p data matrix (denoted X) whose rows represent each of the n cases (in our example, 108 terrorist organizations) and whose columns stand for the p – 1 predictor variables (plus an additional column of 1’s, pertaining to the intercept in the regression equation; p = 13 in our example). Assuming a continuous outcome variable, y (n × 1), matrix notation for the fitted values of y (denoted yˆ ) in the linear regression model is yˆ = X b

(1)

where b is a p × 1 vector of regression coefficients estimated by the ordinary leastsquares criterion. We compute the singular value decomposition (SVD) of X,

X = U S VT

(2)

the point of which (for our purposes) is to produce (as the columns of U) a set of orthogonal dimensions pertaining to the rows of X (the terrorist organizations), and (as columns of V) a set of orthogonal dimensions for the columns of X (the predictor variables). (Superscript T denotes matrix transposition.) S is a diagonal matrix of weights (singular values) indicating relative importance of each dimension. We will be working only with U. Substitution of (2) into (1) shows that eq. (1) is identical to

[UU ] y = yˆ T

(3)

We will refer to the n x n matrix UUT as the projection matrix, following standard usage.2 Eq. (3) is an established result, widely known. We propose to use this result in an innovative way, however. The projection matrix is a network of profile similarity among the cases. In our application, the off-diagonal entries of this matrix indicate the degree of similarity among terrorist organizations i and j with respect to the attributes and behaviors specified in the regression model. We want to turn the regression model “inside out” to express its fitted values for the outcome variable as a function of this network among the cases. 2

This square, symmetric, and idempotent matrix is known in mathematical statistics as the projection matrix, and in the study of regression diagnostics as the hat matrix (e.g., [7]).

Application of a Profile Similarity Methodology

29

Moving on to logistic regression, which is relevant to our application in that CBRN activity is coded as a binary variable, the logistic regression model implies

log(pˆ /1− pˆ ) = X βˆ

(4)

where pˆ is the modeled estimate of the probability that each organization (in turn) engages in CBRN activity, and the vector on the right contains the logistic regression coefficients. We will define a diagonal matrix of weights, W, where wii = pˆ i (1 – pˆ i ). We will compute the SVD of a weighted version of the design matrix X: W 1/2 X = U* S* V*T

(5)

where the superscript asterisks simply indicate that the respective quantities are different from those of eq. (2). Now, although logistic regression is nonlinear, we can transform logistic regression to a form paralleling eq. (3), as follows:

[U* U*T ](W 1/2 z) = (W 1/2 X βˆ )

(6)

where z is a vector of pseudovalues that stand for the observed binary response variable [8].3 The term on the left is the projection matrix for logistic regression [8]. The second term on the left is (a weighted version of) the response variable. The term on the right is (a weighted form of the logits of) the estimated response variable (see (4)). Let’s take a look inside the projection matrix in (6), with respect to Asal and Rethemeyer’s logistic regression analysis of 108 terrorist groups.

4 Results Fig. 1 is a dimensional representation of the cosines among all pairs of rows of matrix U*. This cosine matrix has entries Pij / sqrt (Pii × Pjj) where P = U*U*T is the projection matrix in (6). The dominant eigenvector tends to place organizations with CBRN activity (indicated by dark arrows) on the right, with 13 = al Qaeda at the extreme. Some other organizations are placed with them (e.g., 3 = Abu Sayyaf, 14 = al Qaeda in the Arabian Peninsula, AQAP) and therefore warrant further study. Probabilities of CBRN activity predicted by the logistic regression model are a function of scores on the first eigenvector of Fig. 1. Correlation of those scores with the cube root of the predicted probabilities is +.97 across the 108 cases. The second eigenvector, though not dominant, splits the CBRN-active groups according to differences in their attributes and behaviors. At the top of dimension 2, groups 45, 51, 47, and 50 include three Kashmiri groups (51 = Lashkar-e-Taiba, LeT) and Lashkar-e-Jhangvi (50). Clustering of groups is important because an analyst wanting to identify the smallest set of groups that are “close” in their profiles to the largest number of CBRN-active organizations would do well to select groups from different regions of Fig. 1. For example, consider the top eight groups that each group is most similar to in the projection matrix in eq. (6). Group 51 (LeT) and group 48 (Jemaah Islamiya) are on opposite “sides” of dimension 2 of Fig. 1, and each is 3

The pseudovalues are defined such that exp(z) / (1 + exp(z)) is an excellent approximation of the observed binary response, y (see [8]).

30

R.L. Breiger et al.

Fig. 1. Plot of first two eigenvectors of cosine matrix derived from [U*U*T]. Dark arrows point to organizations coded as using or pursuing CBRN. Organizations are named with numbers, some of which are omitted to enhance clarity.

closest to sets of four CBRN-active groups that do not overlap with each another. Likewise, group 31 (Hamas) is closest to four groups that do not overlap with either of the other two sets of four groups. In aggregate, these three groups (51, 48, 31) share strong profiles with 12 of the 14 CBRN-active groups to be found in our 108group dataset. Clustering matters, because there is not a single profile that characterizes all CBRN-active groups. Clusters or families of groups do not “look like” other clusters. The preceding discussion is important because it reveals clustering (and opportunities to identify distinctively different CBRN profiles) inside the projection matrix that is taken for granted in logistic regression modeling. In order to construct that projection matrix, however, knowledge of the dependent variable is required. (Notice from eq. 5 that matrix W requires the modeled probabilities). In addition to understanding these implications of the logistic regression model, we would like, in the spirit of exploratory analysis, to identify clustering without a reliance on having to know the results of this (or any other) regression model. Therefore, we will now search for clusters relevant to the projection matrix of eq. (3), which depends only on the predictor variables, not on the analyst’s knowledge of the outcome variable. With respect to matrix U of eq. (2), we estimated nine k-means cluster analysis models, yielding two through ten clusters. For each analysis we calculated the sums of squared errors. We then plotted the sum of squares by the number of clusters, yielding a figure that resembles a scree plot in its appearance and interpretation. Based on this plot we determined that a four-cluster solution provides an adequate tradeoff of parsimony vs. fit in defining the clusters for our analyses.

Application of a Profile Similarity Methodology

31

Partitioning the 108 Islamist groups into four clusters yields a 71-group cluster, a 22-group cluster, a 14-group cluster, and a 1-group cluster. The latter cluster (D) consists of an organization (the Islami Chhatra Shibir, ICS) that is an outlier with respect to several of the predictor variables and, consequently, does not have a high profile similarity with any of the other groups. The ICS was founded in 1941 in Bangladesh and is thus more than a quarter-century older than the next oldest group in the dataset. To provide context for the following discussion, we note that the overall baseline probability in this dataset for pursuing or using CBRN weapons is .13 ( = 14 groups out of 108 Islamist groups in the dataset). The largest cluster, with 71 groups (A), contains only one organization coded as using or pursuing CBRN weapons. That group is the East Turkistan Liberation Organization (ETLO), a Uyghur group operating in central Asia. Cluster B, consisting of 22 groups, contains 8 cases coded as pursuing or using CBRN weapons. This is a high concentration (.36) compared with the overall baseline probability of .13. Cluster C contains 14 organizations, of which 5 are coded as using or pursuing CBRN weapons, again a much higher conditional probability (.36) than the overall baseline. Thus, two clusters have conditional probabilities of CBRN activity that are close to 0, and two others have comparatively much higher likelihood. The first column of Table 1 presents means and standard deviations for the variables in our analysis. The remaining columns provide variable means (and standard deviations for quantitative variables) by cluster. In comparison to all groups in the dataset, those groups in cluster B are more connected based on eigenvector centrality (t(106) = 2.4, p0.1, ** p>0.05, *** p>0.01. Standard errors in parentheses. Included in models but not reported are 36 time dummies as control variables.

74

H. Garriga, S. Spaeth, and G. von Krogh

We studied the impact of community mobilization on contributions (Table 2). Model IV shows the positive and significant impact of community mobilization on contributions (P3). One key advantage of public good is usually said to be the involvement of volunteer individuals, and our analysis supports this statement. Since contributions are measured in absolute terms, could a community be expected to halt development? Future research should follow up this topic either through historical methods or simulations. Our study also shows how lower interconnectedness, through modules, translates into increased contributions. Although this was not one of our propositions, it is an intuitive extension of interdependence effects on community mobilization. Table 2. Models explaining contributions through bug reports in OSS public goods Contributions III IV Interconnectedness -1.451** Maturity -0.167 Community mobilization 0.250** Number of committers 0.0348*** 0.025*** Size -9.62e-08*** -3.94e-08** Adj R-sq 0.71 0.77 Statistical significance denoted as: * p>0.1, ** p>0.05, *** p>0.01. Standard errors in parentheses. Included in models but not reported are 36 time dummies as control variables.

As public goods rely on communities’ involvement, it is crucial to have mechanisms that motivate this participation. Interconnectedness impacts community mobilization, as well as contributions to the public good. In Graph 3, we plot the estimated contributions to Eclipse using community mobilization and interconnectedness constructs. We show a trend of increased modules, which partly explains the increased participation. This has previously been proposed [15], but with limited empirical evidence. Our evidence, presented here, contributes to that discussion. However, other factors might add interdependence to the environment, such as concentration [8]. Communities might be reluctant to participate in projects where only one committer has the right to change the repository. Thus, community mobilization might be higher when many developers participate and independency is minimal. Future research should focus on whether the negative correlation between interconnectedness and community mobilization is a general phenomenon for OSS. More studies will be necessary to see whether there is any significant difference in commercial software developments. We suggest that few parallels can be expected, only if they are highly dependent on voluntary contributions. This article contributes to the integration of RDT with collective action theory, and the under-provision and free-rider problem [27]. We look closely at the parameters that hinder the potential of a public good, and, using RDT, explain how provisioning through collective action depends on environmental factors. We examine those factors, which lead to standards of deviance and conflict that alter provisioning [8].

Open Source Software Development: Communities’ Impact on Public Good

75

Graph 3. Contributions through community mobilization and modules in OSS

Fig. 1. Model of public good in OSS development using RDT and collective action

As more actors and resources become part of a public good, synergies can be found that boost contributions, but additional costs and dependencies also arise. This study highlights the importance of interconnectedness for community mobilization, and tentatively underlines its impact on overall contributions, which, however, we leave for future research. It is fundamental to the success of a public good not only to weigh the strengths and weaknesses of the innovation mechanisms available, but also to understand and predict how these mechanisms interact when used in tandem, and to recognize that more is not always better. Future research should consider these new clues, and extend our findings with more empirical evidence and joint theories.

76

H. Garriga, S. Spaeth, and G. von Krogh

References 1. von Hippel, E., von Krogh, G.: Open source software and the ”private-collective” innovation model: Issues for organization science. Organization Science 14(2), 209–223 (2003) 2. Hardin, G.: The Tragedy of the Commons. Science 162(3859), 1243–1248 (1968) 3. Olson, M.: The Logic of Collective Action: Public Goods and the Theory of Groups. Harvard University Press, Cambridge (1965) 4. Osterloh, M., Rota, S.: Open source software development–Just another case of collective invention? Research Policy 36(2), 157–171 (2007) 5. Oliver, P.E., Marwell, G., Teixeira, R.: A theory of the critical mass. I. Interdependence, group heterogeneity, and the production of collective action. American Journal of Sociology 91(3), 522–556 (1985) 6. Marwell, G., Oliver, P.E., Prahl, R.: Social networks and collective action: A theory of the critical mass. III. American Journal of Sociology 94(3), 502–534 (1988) 7. Monge, P.R., et al.: Production of Collective Action in Alliance-Based Interorganizational Communication and Information Systems. Organization Science 9(3), 411–433 (1998) 8. Pfeffer, J., Salancik, G.R.: The external control of organizations: a resource dependence perspective, vol. xiii, p. 300. Harper & Row, New York (1978) 9. Shah, S.: Motivation, governance, and the viability of hybrid forms in open source software development. Management Science 52(7), 1000–1014 (2006) 10. Fulk, J., et al.: Connective and communal public goods in interactive communication systems. Communication Theory 6(1), 60–87 (1996) 11. Aldrich, H.: Resource Dependence and in Terorganiza Tional Relations: Local Employment Service Offices and Social Services Sector Organizations. Administration Society 7(4), 419–454 (1976) 12. Simon, H.A.: The sciences of the artificial. MIT Press, Cambridge (1996), 9780262691918 13. Lerner, J., Tirole, J.: Some simple economics of open source. The Journal of Industrial Economics 50(2), 197–234 (2002) 14. Haefliger, S., von Krogh, G.F., Spaeth, S.: Code Reuse in Open Source Software Development. Management Science 54(1), 180–193 (2008) 15. Baldwin, C.Y., Clark, K.B.: The Power of Modularity. In: Design Rules, vol. 1. The MIT Press, Cambridge (2000) 16. MacCormack, A., Rusnak, J., Baldwin, C.Y.: Exploring the structure of complex software designs: An empirical study of open source and proprietary code. Management Science 52(7), 1015–1030 (2006) 17. West, J., O’Mahony, S.: The Role of Participation Architecture in Growing Sponsored Open Source Communities, pp. 145–168. Carfax Pub. Co., Abingdon (2008) 18. Bechky, B.A.: Sharing meaning across occupational communities: The transformation of understanding on a production floor. Organization Science 14(3), 312–330 (2003) 19. Hars, A., Ou, S.: Working for free? Motivations for participating in open-source projects. International Journal of Electronic Commerce 6(3), 25–39 (2002) 20. Hertel, G., Niedner, S., Herrmann, S.: Motivation of software developers in Open Source projects: an Internet-based survey of contributors to the Linux kernel. Research Policy 32(7), 1159–1177 (2003) 21. von Krogh, G.F., Spaeth, S., Lakhani, K.: Community, joining, and specialization in open source software innovation: a case study. Research Policy 32(7), 1217–1241 (2003)

Open Source Software Development: Communities’ Impact on Public Good

77

22. Bartels, B.: Beyond fixed versus random effects: a framework for improving substantive and statistical analysis of panel, time-series cross-sectional, and multilevel data. The Society for Political Methodology (2008) 23. Roebuck, M.C., Liberman, J.N.: Impact of Pharmacy Benefit Design on Prescription Drug Utilization: A Fixed Effects Analysis of Plan Sponsor Data. Health Services Research 44(3), 988–1009 (2009) 24. Wooldridge, J.M.: Econometric analysis of cross section and panel data. The MIT press, Cambridge (2002) 25. Pesaran, M.H.: General diagnostic tests for cross section dependence in panels (2004) 26. Hoechle, D.: Robust standard errors for panel regressions with cross-sectional dependence. Stata Journal 7(3), 281 (2007) 27. Ostrom, E.: Governing the commons: The evolution of institutions for collective action. Cambridge Univ. Pr., Cambridge (1990)

Location Privacy Protection on Social Networks Justin Zhan and Xing Fang Department of Computer Science North Carolina A&T State University {Zzhan,xfang}@ncat.edu

Abstract. Location information is considered as private in many scenarios. Protecting location information on mobile ad-hoc networks has attracted much research in past years. However, location information protection on social networks has not been paid much attention. In this paper, we present a novel location privacy protection approach on the basis of user messages in social networks. Our approach grants flexibility to users by offering them multiple protecting options. To the best of our knowledge, this is the first attempt to protect social network users’ location information via text messages. We propose five algorithms for location privacy protection on social networks. Keywords: Location information; privacy; social networks.

1 Introduction Online social network has gained remarkable popularity since its debut. One of the most popular online social networks, Facebook, has reached more than 500 million active users [4]. Social network users are able to post their words, pictures, videos, etc. They can also write comments, share personal preferences, and make friends. Facebook even integrates messenger on their webpages supporting users instant online chatting with current online friends. Online social networks typically deal with large amounts of user data. This may lead to privacy-related user data revelation [3, 14, 16]. Personal information, such as personal interests, contact information, photos, activities, associations and interactions, once revealed, may have different levels of impact, ranging from unexpected embarrassment or reputational damage [11] to identity theft [15]. Furthermore, effectively managing privacy for social network can be quite tricky. One reason is that different individuals have different levels of privacy-related expectations towards their information. For instance, some of them may be glad to publish their personal profiles tending to potentially develop additional friendships, while others may worry about the exposure of their identities so as to only reveal the profiles to their selected friends. In other situations, people are even willing to disclose their personal information to anonymous strangers rather than acquaintances [6]. Unfortunately, people may not often care about their privacy. Gross and Acquisti [6] found that only a few students change the default privacy preferences on Facebook by analyzing and evaluating the online behavior as well as the amount of private information disclosed from 4000 students in Carnegie Mellon University. Location information may be J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, pp. 78–85, 2011. © Springer-Verlag Berlin Heidelberg 2011

Location Privacy Protection on Social Networks

79

sensitive because it enables the possibility of tracing someone. Motivated by this reason, we propose our privacy location information protection algorithms using encryption, k-anonymity and noise injection techniques. The rest of this paper is organized as follows: In section 2, we briefly review some related work on location privacy protection. In section 3, we introduce our location privacy protection approach. We conduct experimental evaluation and present the result in section 4. We conclude the paper in section 6.

2 Related Work 2.1 Location Privacy Protection on Mobile Networks Location privacy protection has gained much attention on mobile networks. With the rapid development of computing technology, computational capability harnessed from mobile phone hardware enables the running of some powerful software applications. Location-based service (LBS) is one of the applications where its service provider requires users to provide their location information in order to response their requests [2]. Schilit et al. [12] pointed out that location-based service is prone to privacy revelation. The service even incurs economic and reputation damages because of its potential privacy risks. Amoli et al. [2] classified privacy protections on LBS into two categories: policy-based approaches and data hiding techniques. In the policy-based approaches, a set of pre-defined policies are applied to users to inform that when and how the data can be stored, used, and revealed. K-anonymity, dummy-based approach, and obfuscation are three methods for the data hiding techniques. The kanonymity works when a user’s location can be hidden within a set of k members, who possess the same location information. In the dummy-based approach, a user is able to create some dummy location information to the service provider instead of sending out the real location information, while the obfuscation allows users to generalize their actual locations before sending them out. Since the aforementioned approaches have shortcomings, Amoli et al. proposed 2Ploc, a privacy preserving protocol of LBS based on one-time tickets [2]. Location protections also apply to mobile ad hoc routing protocols. Kamat et al. [8] introduced an idea of using fake routing sources. The fake sources are able to lead adversaries to fake locations. Kong and Hong [9] proposed ANDOR, an anonymous routing protocol, which suggests using route pseudonyms instead of node IDs during the routing process. ARM [13] is another anonymous routing protocol in which two nodes share a secret key and a pseudonym. Based on ANDOR and ARM, Taheri et al. [18] further proposed RDIS, an anonymous routing protocol with destination location privacy protection. 2.2 Location Privacy Protection on Social Networks Location information on social networks is commonly treated as private data and is only available to a certain group of people. For instance, users of online social networks are able to configure their privacy settings so as to only reveal their locations to friends. Unfortunately, the majority of social network users only make changes of their default privacy settings after bad things occurred and the existing privacy settings are perplexed. Lipford et al. [10] presented a new privacy setting

80

J. Zhan and X. Fang

interface based on Facebook that makes significantly improved understanding on the settings as well as better performance. The interface enables a set of HTML tabs, each providing a different browser’s view of Facebook users’ account information, in order to let the users properly manage their privacy settings. Ho et al. [7] analyzed the most popular social networks and discovered three privacy problems. First, users are not notified by social networks when their personal information is at privacy risks. Second, existing privacy protection tools in social networks are not flexible enough. Third, users cannot prevent information that may reveal the privacy of themselves from being uploaded by any other users. To solve the problems, Ho et al. [7] designed a privacy management framework, in which it enables data levels, privacy levels, and tracking levels, respectively. By designating data levels, users are able to sort their data into different privacy-related levels. Different amount of user data is available for different viewers, such as visitors, casual friends, normal friends, and best friends, according to their privacy levels. Finally, tracking levels can be applied to let users decide the magnitude of their account information that can be tracked. Content on social networks is usually tagged by its poster. This incurs privacy threats when location-related content is tagged via spatial-temporal tags. Freni et al. [5] summarized such threats as the location privacy and the absence privacy. The former one enables the possibility of referring other users’ presence of specific locations at a given period of time. The latter one enables the possibility of referring other users’ absence of specific locations at a given period of time. Freni et al. then introduced the generalization methods to prevent the privacy threats, where the methods are primarily space-based generalization and time-based generalization. Although all the aforementioned location privacy protection schemes have their advantages, to some extent, they are not applicable to protect location information revelation on user text messages. In this paper, we propose our location protection approach based on user text messages.

3 Our Approach 3.1 Preliminaries Social network users frequently share text information with others. Abrol and Khan [1] presented a location information extraction approach on the basis of users’ messages. In their approach, location information of a certain message is tagged and ranked based on its collision with their geo-location database entries. The one with the highest ranking score will be extracted. In our approach, we maintain a similar geo-location database, for the purpose of verifying location information’s existence for a given user massage. We narrow down our geo-location information to city level in our database. Hence, for each entry of data, it includes city, zip code, state, and country. The three location information protection techniques provided in our approach are encryption, k-anonymity, and noise injection. Users are able to make choice of them based on their concerns. Encryption is recommended when a user wants to share her location information with designated individuals, who are the friends of the user. In this case, the entire user message will be encrypted by its receiver’s public key. Both k-anonymity and noise injection techniques are suggested when a user wants to publicly post her message, in which it incorporates her location information.

Location Privacy Protection on Social Networks

81

In the rest of this subsection, we will formally elaborate some notions. Our geolocation database, DB, is a universe of location information of United States, specified to city level. Each data entry resembles as {city, zip code, state, country}. stands for a user that the user’s location information is tagged as . A user message is defined as , where it is tagged as . By denoting , we claim that | 0, we claim that contains user location information. By denoting | the tagged location information matches the database entries. For every user, she maintains a list of friends’ public keys . 3.2 Encryption Approach The encryption technique applies when a user wants to share her location information with her friends. Algorithm 1 is the pseudo code of the encryption algorithm. Algorithm 1. The Encryption Approach Input: Output: 1: for every do 2: 3: end for 4: if 5: return 6: else if (Encrypt = =1and 7: 8: else 9: return 10:end if

then ) then

In the algorithm, it takes user message as its input and outputs the encrypted cyphertext, which is able to be decrypted by its receiver’s private key. Location information of is tagged by the function. The Encrypt is a Boolean variable. When its value equals one, it means the user enables the encryption technique. Otherwise, the encryption technique is disabled. This leaves flexibility to user when choosing location protection techniques. 3.3 K-anonymity Approach K-anonymity technique is recommended when a user wants to publicly share her message, in which the message incorporates the user’s location information. Algorithm 2. The k-anonymity Approach Input: Output: 1: for every 2: 3: end for 4: if

do then

82

J. Zhan and X. Fang

5: return 6: else if (ka = =1) then 7: 8: end if 9: return Algorithm 3. The k-anonymize algorithm Input: , Output: 1:for every 2: if (| 3: return 4: else 5: return 6: end if 7: end for 8: 9: return

|

do 2) then

Each data entry, in the , resembles as {city, zip code, state, country}. Since we only consider the geo-location information throughout United States, the data entry is able to be reduced to {city, zip code, state}, without losing any information. A totality , , , . tagging of a certain data entry resembles as , , It is widely known that a zip code is a unique identifier for a city. Therefore, the code in line 4 of the k-anonymity function is able to remove zip code tag for a given . The collection of tags then becomes , , , . The tag, , , is a tagging of the Quasi-identifier of our . And this tag is able to be, at least, 2-anonymized via suppression, after the 4th step of the k-anonymity function. According to Sweeney [17], k-anonymity of a given message’s location information is then able to be achieved as long as the Quasi-identifier of the location information is k-anonymized. Given an example that the information, “Madison, SD, 57042”, is tagged from a . The k-anonymity algorithm first suppresses the zip code tag. The message transfers to “Madison, SD, *****”. The Quasi-identifier tag then still uniquely matches our database entry, because there is only one city named Madison in South Dakota. Therefore, two ways of suppressions can be applied to anonymize the Quasi-identifier tag. One is to completely anonymize the tag, which it leads to “*******, **, *****”. The alternative way is to anonymize the state tag in order to achieve city level anonymity. 3.4 Noise Injection Approach The noise injection technique is aimed to protect user location information by injecting location noise, , in between . City level and state level anonymity can be achieved via noise injection. In the noise injection algorithm, we still leave user to decide which level’s anonymity is most appropriate to her.

Location Privacy Protection on Social Networks

83

Algorithm 4. The Noise Injection Approach Input: Output: 1: for every do 2: 3: end for 4: if 5: return 6: else if (ni = =1) then 7: 8: end if 9: return

then

Algorithm 5. The Noise-Inject algorithm Input: , Output: 1: if 2: for every 3: 4: end for 5: else if 6: for every 7: 8: end for 9: for every

1 then do , 1 then do do

10: 11: end for 12:else return 13:end if 14: 15: return | 2. Consider a Recall that k-anonymity is achieved when | , , and noise , ,a with injected noise is denoted as || , , || . Therefore, a noise injected matches at least two entries of the . Since noise city and zip code both belong to the same state as the original city and zip code, city level kinferred anonymity is accomplished that the original city is hidden with 1 possibility. For the state level, first, a ′ city and zip code tags are suppressed that only leaves state tag unsuppressed. Hence, state level k-anonymity is able to accomplish via injecting, at least, one noisy state. Similarly, the state level kanonymity has 1 inferred possibility.

84

J. Zhan and X. Fang

4 Conclusion Location privacy protection is an important research topic on mobile ad-hoc networks. However, location privacy protection on social networks has not been well studied. In this paper, we present our approach for the protection of location information based on social network user text messages. Three techniques including encryption, k-anonymity, and noise injection, are provided. We grant the right of selecting the techniques by users. Our experimental results show that our approach can significantly reduce location privacy revelation and can be efficiently deployed. Even if our location protection approach is able to well performed, there are still some limitations haunting on it. First, the geo-location database has shortcomings. Currently, we only collected city level geo-locations in United States. However, social network users are scattered around the world, which requires us to expand the database entries in order to protect users outside of U.S. Additionally, we also need to expand our database to include any other geo-location information lower than city level. For instance, if a user message mentions locations like district, street, or buildings, our current database is then not able to deal with this situation. Second, our approach grants users the flexibilities for selecting protection techniques. On the other hand, it also impose burden such as the understanding of protection techniques to users. As our future work, we will integrate linguistic analysis towards user messages to allow system to automatically select the best protection technique for each user.

References 1. Abrol, S., Khan, L.: TweetHood: Agglomerative Clustering on Fuzzy k-Closest Friends with Variable Depth for Location Mining. In: Proceedings of the IEEE International Conference on Privacy, Security, Risk and Trust, Minneapolis, MN, USA, pp. 153–160 (August 2010) 2. Amoli, A., Kharrazi, M., Jalili, R.: 2Ploc: Preserving Privacy in Location-Based Services. In: Proceedings of the IEEE International Conference on Privacy, Security, Risk and Trust, Minneapolis, MN, USA, pp. 707–712 (August 2010) 3. Blakely, R.: Does Facebook’s Privacy Policy Stack Up?, http://business.timesonline.co.uk/tol/business/ industry_sectors/technology/article2430927.ece 4. Facebook Statistics, http://www.facebook.com/press/info.php?statistics 5. Freni, D., Vicente, C., Mascetti, S., Bettini, C., Jensen, C.: Preserving Location and Absence Privacy in Geo-Social Networks. In: Proceedings of the 19th ACM International Conference on Information and Knowledge Management, Toronto, Ontario, Canada (October 2010) 6. Gross, R., Acquisti, A.: Information Revelation and Privacy in Online Social Networks. In: Proceedings of the 2005 ACM Workshop on Privacy in the Electronic Society, Alexandria, Virginia, USA, pp. 71–80 (November 2005) 7. Ho, A., Maiga, A., Aimeur, E.: Privacy Protection Issues in Social Networking Sites. In: Proceedings of IEEE/ACS International Conference on Computer Systems and Applications, Rabat, Morocco, pp. 271–278 (May 2009)

Location Privacy Protection on Social Networks

85

8. Kamat, P., Zhang, Y., Trappe, W., Ozturk, C.: Enhancing Source-Location Privacy in Sensor Network Routing. In: Proceedings of the 25th IEEE International Conference on Distributed Computing Systems, Columbus, Ohio, USA (June 2005) 9. Kong, J., Hong, X.: ANODR: Anonymous on Demand Routing with Untraceable Routes for Mobile ad-hoc Networks. In: Proceedings of the 4th ACM International Symposium on Mobile ad hoc Networking and Computing, New York, NY, USA, June 2003, pp. 291–302 (2003) 10. Lipford, H., Besmer, A., Watson, J.: Understanding Privacy Settings in Facebook with an Audience View. In: Proceedings of the 1st Conference on Usability, Psychology, and Security, San Francisco, USA, pp. 1–8 (April 2008) 11. Rosenblum, D.: What Anyone Can Know: The Privacy Risks of Social Networking Sites. IEEE Security and Privacy 5(3), 40–49 (2007) 12. Schilit, B., Hong, J., Gruteser, M.: Wireless Location Privacy Protection. Computer 36(12), 135–137 (2003) 13. Seys, S., Preneel, B.: ARM: Anonymous Routing Protocol for Mobile ad hoc Networks. International Journal of Wireless and Mobile Computing 3(3) (October 2009) 14. Sogholan, C.: The Next Facebook Privacy Scandal, http://news.cnet.com/8301-13739_3-9854409-46.html 15. Strater, K., Lipford, H.: Strategies and Struggles with Privacy in an Online Social Networking Community. In: Proceedings of the 22nd British HCI Group Annual Conference, Liverpool, UK, pp. 111–119 (September 2008) 16. Stross, R.: When Everyone’s a Friend, Is Anything Private?, http://www.nytimes.com/2009/03/09/business/worldbusiness/ 09iht-08digi.20688637.html 17. Sweeney, L.: k-anonymity: A Model for Protecting Privacy. International Journal on Uncertainty, Fuzziness and Knowledge-based Systems 10(5), 557–570 (2002) 18. Taheri, S., Hartung, S., Hogrefe, D.: Achieving Receiver Location Privacy in Mobile ad hoc Networks. In: Proceedings of the IEEE International Conference on Privacy, Security, Risk and Trust, Minneapolis, MN, USA, pp. 800–807 (August 2010)

Agent-Based Models of Complex Dynamical Systems of Market Exchange Herbert Gintis The Santa Fe Institute and Central European University [email protected]

Abstract. The standard Walrasian general equilibrium model is a static description of market clearing equilibria. Attempts over more than a century to provide a decentralized market dynamic that implements market equilibrium have failed. The reason is that a system of decentralized markets is a complex dynamical system in which the major form of learning is through experience (adaptive expectations) and imitating successful others (replicator dynamics). I will present a model of decentralized market exchange where each individual produces one good and consumes many. I will show that an agentbased model of this economy converges to market equilibrium and is highly impervious to shocks. I will also present a model in which agents are firms and households, as in the standard Walrasian model. I will show that an agent-based model in this case converges strongly to market equilibrium, but with considerable excess volatility and large excursions from equilibrium. This is characteristic of error terms with "fat tails" as characterized in the complexity literature, and is due to the tendency of agents to imitate the successful, which leads to strongly correlated error distributions.

J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, p. 86, 2011. © Springer-Verlag Berlin Heidelberg 2011

Computational and Statistical Models: A Comparison for Policy Modeling of Childhood Obesity PANEL MEMBERS Chair/Moderator: Patricia L. Mabry, Ph.D. Senior Advisor Office of Behavioral and Social Sciences Research Office of the Director National Institutes of Health Participants: Ross Hammond, Ph.D. Director, Center on Social Dynamics & Policy Senior Fellow, Economic Studies Brookings Institution

Terry T-K Huang, PhD, MPH Professor and Chair Department of Health Promotion & Social and Behavioral Health College of Public Health University of Nebraska Medical Center

Edward Hak-Sing Ip , Ph.D. Professor and Associate Chair of Research & Faculty Development Division of Public Health Sciences Wake Forest University Health Sciences Abstract. As systems science methodologies have begun to emerge as a set of innovative approaches to address complex problems in behavioral, social science, and public health research, some apparent conflicts with traditional statistical methodologies for public health have arisen. Computational modeling is an approach set in context that integrates diverse sources of data to test the plausibility of working hypotheses and to elicit novel ones. Statistical models are reductionist approaches geared towards proving the null hypothesis. While these two approaches may seem contrary to each other, we propose that they are in fact complementary and can be used jointly to advance solutions to complex problems. Outputs from statistical models can be fed into computational models, and outputs from computational models can lead to further empirical data collection and statistical models. Together, this presents an iterative process that refines the models and contributes to a greater understanding of the problem and its potential solutions. The purpose of this panel is to foster communication and understanding between statistical and computational modelers. Our goal is to shed light on the differences between the approaches and convey what kinds of research inquiries each one is best for addressing and how they can serve complementary (and synergistic) roles in the research process, to mutual benefit. For each J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, pp. 87–88, 2011. © Springer-Verlag Berlin Heidelberg 2011

88

P.L. Mabry et al.

approach the panel will cover the relevant “assumptions” and how the differences in what is assumed can foster misunderstandings. The interpretations of the results from each approach will be compared and contrasted and the limitations for each approach will be delineated. We will use illustrative examples from CompMod, the Comparative Modeling Network for Childhood Obesity Policy. The panel will also incorporate interactive discussions with the audience on the issues raised here.

Detecting Changes in Opinion Value Distribution for Voter Model Kazumi Saito1 , Masahiro Kimura2 , Kouzou Ohara3 , and Hiroshi Motoda4 1

3

School of Administration and Informatics, University of Shizuoka       2 Department of Electronics and Informatics, Ryukoku University      Department of Integrated Information Technology, Aoyama Gakuin University   4 Institute of Scientific and Industrial Research, Osaka University    

Abstract. We address the problem of detecting the change in opinion share over a social network caused by an unknown external situation change under the valueweighted voter model with multiple opinions in a retrospective setting. The unknown change is treated as a change in the value of an opinion which is a model parameter, and the problem is reduced to detecting this change and its magnitude from the observed opinion share di usion data. We solved this problem by iteratively maximizing the likelihood of generating the observed opinion share, and in doing so we devised a very eÆcient search algorithm which avoids parameter value optimization during the search. We tested the performance using the structures of four real world networks and confirmed that the algorithm can eÆciently identify the change and outperforms the naive method, in which an exhaustive search is deployed, both in terms of accuracy and computation time. Keywords: Voter Model, Opinion Share, Change Point Detection.

1 Introduction Recent technological innovation in the web such as blogosphere and knowledge mediasharing sites is remarkable, which has made it possible to form various kinds of large social networks, through which behaviors, ideas and opinions can spread, and our behavioral patterns are strongly aected by the interaction with these networks. Thus, substantial attention has been directed to investigating the spread of influence in these networks [9,2,14]. Much of the work has treated information as one entity and nodes in the network are either active (influenced) or inactive (uninfluenced), i.e. there are only two states. However, application such as an on-line competitive service in which a user can choose one from multiple choices decisions requires a model that handles multiple states. In addition, it is important to consider the value of each choice, e.g., quality, brand, authority, etc. because this impacts our choice. We formulated this problem using a value-weighted K opinion diusion model and provided a way to accurately predict J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, pp. 89–96, 2011. c Springer-Verlag Berlin Heidelberg 2011

­

90

K. Saito et al.

the expected share of the opinions at a future target time from a limited amount of observed data [6]. This model is an extension of the basic voter model which is based on the assumption that a person changes its opinion by the opinions of its neighbors. There has been a variety of work on the voter model. Dynamical properties of the basic model have been extensively studied including how the degree distribution and the network size aect the mean time to reach consensus [10,12]. Several variants of the voter model are also investigated and non equilibrium phase transition is analyzed [1,15]. Yet another line of work extends the voter model by combining it with a network evolution model [3,2]. These studies are dierent from what we address in this paper. Almost all of the work so far on information diusion assumed that the model is stationary. However, our behavior is aected not only by the behaviour of our neighbors but also by other external factors. We apply our voter model to detect a change in opinion share which is caused by an unknown external situation change. We model the change in the external factors as a change in the opinion value, and try to detect the change from the observed opinion share diusion data. If this is possible, this would bring a substantial advantage. We can detect that something unusual happened during a particular period of time by simply analyzing the data. Note that our approach is retrospective, i.e. we are not predicting the future, but we are trying to understand the phenomena that happened in the past, which shares the same spirit of the work by Kleinberg [7] and Swan [13] in which they tried to organize a huge volume of the data stream and extract structures behind it. Thus, our problem is reduced to detecting where in time and how long this change persisted and how big this change is. To make the analysis simple, we limit the form of the value change to a rect-linear one, that is, the value changes to a new higher level, persists for a certain period of time and is restored back to the original level and stays the same thereafter. We call this period when the value is high as “hot span” and the rest as “normal span”. We use the same parameter optimization algorithm as in [6], i.e. the parameter update algorithm based on the Newton method which globally maximizes the likelihood of generating the observed data sequences. The problem here is more diÆcult because it has another loop to search for the hot span on top of the above loop. The naive learning algorithm has to iteratively update the patten boundaries (outer loop) and the value must also be optimized for each combination of the pattern boundaries (inner loop), which is extraordinary ineÆcient. We devised a very eÆcient search algorithm which avoids the inner loop optimization during the search. We tested the performance using the structures of four real world networks (blog, Wikipedia, Enron and coauthorship), and confirmed that the algorithm can eÆciently identify the hot span correctly as well as the opinion value. We further compared our algorithm with the naive method that finds the best combination of change boundaries by an exhaustive search through a set of randomly selected boundary candidates, and showed that the proposed algorithm far outperforms the native method both in terms of accuracy and computation time.

2 Opinion Formation Models The mathematical model we use for the diusion of opinions is the value-weighted voter model with K ( 2) opinions [6]. A social network is represented by an undirected

EÆcient Detection of Hot Span in Information Di usion

91

(bidirectional) graph with self-loops, G  (V E), where V and E ( V  V) are the sets of all the nodes and links in the network, respectively. For a node v  V, let  (v) denote the set of neighbors of v in G, that is,  (v)  u  V; (u v)  E . Note that v   (v). In the model, each node of G is endowed with (K  1) states; opinions 1,    , K, and neutral (i.e., no-opinion state). It is assumed that a node never switches its state from any opinion k to neutral. The model has a parameter wk ( 0) for each opinion k, which is called the value-parameter and must be estimated from observed opinion diusion data. Let ft : V  0 1 2    K  denote the opinion distribution at time t, where ft (v) stands for the opinion of node v at time t, and opinion 0 denotes the neutral state. We also denote by nk (t v) the number of v’s neighbors that hold opinion k at time t for k  1 2    K, i.e., nk (t v)  u   (v); ft (u)  k . Given a target time T , and an initial state in which each opinion is assigned to only one distinct node and all other nodes are in the neutral state, the evolution process of the model unfolds in the following way. In general, each node v considers changing its opinion based on the current opinions of its neighbors at its ( j 1)th update-time t j 1 (v), and actually changes its opinion at the jth update-time t j (v), where t j 1 (v)  t j (v)  T , j  1 2 3    , and t0 (v)  0. It is noted that since node v is included in its neighbors by definition, its own opinion is also reflected. The jth update-time t j (v) is decided at time t j 1 (v) according to the exponential distribution of parameter  (we simply use   1 for any v  V)1 . Then, node v changes its opinion at time t j (v) as follows: If node v has at least one neighbor with some opinion at time t j 1 (v), ft j (v) (v)  k with probability wk nk (t j 1 (v) v) K  k 1 wk nk (t j 1 (v) v) for k  1    K, otherwise, ft j (v) (v)  0 with probability 1. Note here that ft (v)  ft j 1 (v) (v) for t j 1 (v)  t  t j (v). If the next update-time t j (v) passes T , that is, t j (v)  T , then the opinion evolution of v is over. The evolution process terminates when the opinion evolution of every node in G is over. Given the observed opinion diusion data (T s T e )  (v t ft (v)) in time-interval [T s T e ] (a single example), we consider estimating the values of value-parameters w1 ,    , wK , where 0  T s  T e  T . From the evolution process of the model, we can obtain the following log likelihood function ¼

¼

¼

 

(w; (T s

T e ))  log



(v t k)(T s T e )

nk (t v)wk K k¼ 1 nk¼ (t v)wk¼

(1)

where w  (w1    wK ) stands for the K-dimensional vector of value-parameters, and (T s T e )  (v t ft (v))  (T s T e ); u   (v); ft (u)  0 2. Thus, our estimation problem is formulated as a maximization problem of the objective function

(w; (T s T e )) with respect to w. We find the optimal values of w by employing a standard Newton method (see [6] for more details).

3 Change Detection Problem We investigate the problem of detecting the change in behavior of opinion diusion in a social network G based on the value-weighted voter model with K opinions, which is referred to as the change detection problem. In this problem, we assume that some 1

Note that this is equivalent to picking a node randomly and updating its opinion in turn V times.

92

K. Saito et al.

change has happened in the way the opinions diuse, and we observe the opinion diffusion data in which the change is embedded, and consider detecting where in time and how long this change persisted and how big this change is. Here, we mathematically formulate the change detection problem. For the opinion diusion data (0 T ) in time-interval [0 T ], let [T 1 T 2 ] denote the hot (change) span of the diusion of opinions. This implies that the intervals [0 T 1 ) and (T 2 T ] are the normal spans. Let wn and wh denote the value-parameter vectors for the normal span and the hot span, respectively. Note that wn  wn  wh  wh since the opinion dynamics under the value-weighted voter model is invariant to positive scaling of the value-parameter vector w, where wn  and wh  stand for the norm of vectors wn and wh . Then, the change detection problem is formulated as follows: Given the opinion diusion data (0 T ) in time-interval [0 T ], detect the anomalous span [T 1 T 2 ], and estimate the value-parameter vector wh of the hot span and the value-parameter vector wn of the normal span. Since the value-weighted voter model is a stochastic process model, every sample of opinion diusion can behave dierently. This means that it is quite diÆcult to accurately detect the true hot span from only a single sample of opinion diusion. Methods that use only the observed bursty activities, including those proposed by Swan and Allan [13] and Kleinberg [7] would not work. We believe that an explicit use of underlying opinion diusion model is essential to solve this problem. It is crucially important to detect the hot span precisely in order to identify the external factors which caused the behavioral changes.

4 Detection Methods 4.1 Naive Method Let   t1    tN  be a set of opinion change time points of all the nodes appearing in the diusion results (0 T ). We can consider the following value-parameter vector switching when there is a hot span S  [T 1 T 2 ]: w

w

n if t wh if t

 S   S



Then, an extended objective function (wn wh ; (0 T ) S ) can be defined by adequately modifying Equation (1) under this switching scheme. Clearly, the extended objective function is expected to be maximized by setting S to be the true span S   [T 1 T 2 ], for which (0 T ) is generated by the value-weighted voter model, provided that (0 T ) is suÆciently large. Therefore, our hot span detection problem is formalized as the following maximization problem. Sˆ



arg max (w ˆ n wˆ h ; (0 T ) S ) S

(2)

where w ˆ n and wˆ h denote the maximum likelihood estimators for a given S . In order to obtain Sˆ according to Equation (2), we need to prepare a reasonable set of candidate spans, denoted by . One way of doing so is to construct  by considering all pairs of observed activation time points. Then, we can construct a set of candidate

EÆcient Detection of Hot Span in Information Di usion

93

spans by   S  [t1 t2 ] : t1  t2 t1   t2   . Equation (2) can be solved by a naive method which has two iterative loops. In the inner loop we first obtain the maximum likelihood estimators, wˆ n and wˆ h , for each candidate S by maximizing the objective function (wn wh ; (0 T ) S ) using the Newton method. In the outer loop we select the optimal Sˆ which gives the largest (w ˆ n wˆ h ; (0 T ) S ) value. However, this method can be extremely ineÆcient when the number of candidate spans is large. Thus, in order to make it work with a reasonable computational cost, we consider restricting the number of candidate time points to a small value, denoted by J, i.e., we construct J  t1    t J  by selecting J points from  ; then we construct a restricted set of candidate spans by  J  S  [t1 t2 ] : t1  t2 t1  J t2   J . Note that  J  J(J 1)2, which is large when J is large. 4.2 Proposed Method It is easily conceivable that the naive method can detect the hot span with a reasonably good accuracy when we set J large at the expense of the computational cost, but the accuracy becomes poorer when we set J smaller to reduce the computational load. We propose a novel detection method below which alleviates this problem and can eÆciently and stably detect a hot span from diusion results (0 T ). We first obtain the maximum likelihood estimators, w ˆ based on the original objective function of Equation (1), and focus on the first-order derivative of the objective function

(w; (0 T )) with respect to the value-parameter vector w at each individual opinion change time. More specifically, let wt be the value-parameter vector at time t   . Then we obtain the following formula for the maximum likelihood estimators due to the uniform parameter setting and the globally optimal condition.

(w; ˆ (0





T ))



w





(w; ˆ (0

T ))

wt





t

0

Now, we can consider the following partial sum for a given hot span S g(S ) 



(w; ˆ (0





t S

wt

T ))





[T 1 T 2 ]. (4)



Clearly,  g(S ) is likely to have a suÆciently large positive value if S problem setting. Namely, the hot span is detected as follows: Sˆ

(3)

arg max  g(S )

 S  due to our (5)

S 

Here note that we can incrementally calculate g(S ). More specifically, let   tN  be a set of candidate time points, where ti  t j if i  j; then, we can obtain the following formula.

t1   

g([ti t j1 ])  g([ti t j ]) 

(w; ˆ (0





wt j1

T ))



(6)

The computational cost of the proposed method for examining each candidate span is much smaller than the naive method described above. When   N is very large, we construct a restricted set of candidate spans J as explained above. We summarize our proposed method below.

94

1. 2. 3. 4.

K. Saito et al.

Maximize (w; (0 T )) by using the Newton method. Construct the candidate time set  and the candidate span set . Detect a hot span Sˆ by Equation (5) and output Sˆ . Maximize (wn wh ; (0 T ) Sˆ ) by using the Newton method, and output (wˆn wˆh ).

Here note that the proposed method requires likelihood maximization by using the Newton method only twice.

5 Experimental Evaluation We adopted four datasets of large real networks. They are all bidirectionally connected networks. The first one is a trackback network of Japanese blogs used in [5], which has 12 047 nodes and 79 920 directed links (the blog network). The second one is a network of people that was derived from the “list of people” within Japanese Wikipedia, used in [4], and has 9 481 nodes and 245 044 directed links (the Wikipedia network). The third one is a network derived from the Enron Email Dataset [8] by extracting the senders and the recipients and linking those that had bidirectional communications. It has 4 254 nodes and 44 314 directed links (the Enron network). The fourth one is a coauthorship network used in [11], which has 12 357 nodes and 38 896 directed links (the coauthorship network). For each of these networks, we generated opinion diusion results for three dierent values of K (the number of opinions), i.e., K  2, 4, and 8, by choosing the top K nodes with respect to node degree ranking as the initial K nodes and simulating the model mentioned in section 2 from 0 to T  25. We assumed that the value of all the opinions were initially 10, i.e. the value-parameters for all the opinions are 10 for the normal span, and further assumed that the value of the first opinion changed to double for a period of [10 15], i.e. the value-parameter of the first opinion is 20 and the value-parameters of all the other opinions are 10 for the hot span. We then estimated the hot span and the value-parameters for both the spans (normal and hot) by the two methods (the proposed and the naive), and compared their accuracy and the computation time. We adopted 1 000 as the value of J (the number of candidate time points) for the proposed method, and 5, 10, and 20 for the naive method. Figures 1 and 2 show the experimental results2 where each value is the average over 10 trials for 10 distinct diusion results. We evaluated the accuracy of the estimated hot span [Tˆ 1 Tˆ 2 ] by the absolute error Tˆ 1 T 1  Tˆ 2 T 2 , and the accuracy of the estimated opinion values w ˆ by the mean absolute error iK 1 ( wˆ in win  wˆ ih wih )K, where win and wih are values of opinion i for the normal and the hot spans, respectively. From these results, we can find that the proposed method is much more accurate than the naive method for both the networks. The average error for the naive method decreases as J becomes larger. But, even in the best case for the naive method (J  20), its average error in the estimation of the hot span is maximum about 30 times larger than that of proposed method (in the case of the Enron network under K  2), and it is maximum about 6 times larger in the estimation of value-parameters (in the case of the coauthorship network under K  2). It is noted that the naive method needs much 2

We only show the results for the two networks (Enron and coauthorship) due to the space limitation. In fact, we obtained similar results also for the other two networks (blog and Wikipedia).

EÆcient Detection of Hot Span in Information Di usion 4.5

4.8

2.5

5.0

1.0 1.0

0.0 0.0

0.2 0.2

2.2

2

4 #Opinion K

0.1 0.1

0.0 0.0

8

Proposed method Naive J = 5 Naive J = 10 Naive J = 20

(a) Error of [Tˆ 1 Tˆ 2 ]

2

4 #Opinion K

80.3

30.0 30.0

Computation time (s)

8.1

Error

Error

2.0 2.0

95

15.0 15.0

0.0 0.0

8

(b) Error of wˆ

2

4 #Opinion K

8

(c) Computation time

Fig. 1. Comparison on the Enron network 2.2

4.3

4.8

1.0 1.0

0.0 0.0

0.2 0.2

2

4 #Opinion K

8

(a) Error of [Tˆ 1 Tˆ 2 ]

Proposed method Naive J = 5 Naive J = 10 Naive J = 20

0.1 0.1

0.0 0.0

2

4 #Opinion K

(b) Error of wˆ

8

57.4

30.0 30.0

Computation time (s)

3.8

Error

Error

2.0 2.0

15.0 15.0

0.0 0.0

2

4 #Opinion K

8

(c) Computation time

Fig. 2. Comparison on the coauthorship network

longer computation time to achieve these best accuracies than the proposed method although the number of candidate time points for the naive method is 50 times smaller. Indeed, it is about 20 times longer for the former case, about 13 times longer for the latter case, and maximum about 95 times longer for the whole results (in the case of the Enron network under K  8). From these results, it can be concluded that the proposed method is able to detect and estimate the hot span and value-parameters much more accurately and eÆciently compared with the naive method.

6 Conclusions In this paper, we addressed the problem of detecting the unusual change in opinion share from the observed data in a retrospective setting, assuming that the opinion share evolves by the value-weighted voter model with multiple opinions. We defined the hot span as the period during which the value of an opinion is changed to a higher value than the other periods which are defined as the normal spans. A naive method to detect such a hot span would iteratively update the pattern boundaries that form a hot span (outer loop) and iteratively update the opinion value for each hot span candidate (inner loop) such that the likelihood function is maximized. This is very ineÆcient and totally unacceptable. We developed a novel method that avoids the inner loop optimization during search. It only needs to estimate the value twice by the iterative updating algorithm (Newton method), which can reduce the computation times by 7 to 95 times, and is very eÆcient. We applied the proposed method to opinion share samples generated from four real world large networks and compared the performance with the naive method that considers only the randomly selected boundary candidates. The results clearly indicate that the proposed method far outperforms the naive method both

96

K. Saito et al.

in terms of accuracy and eÆciency. Although we assumed a simplified problem setting in this paper, the proposed method can be easily extended to solve more intricate problems. As the future work, we plan to extend this framework to spatio-temporal hot span detection problems.

Acknowledgments This work was partly supported by Asian OÆce of Aerospace Research and Development, Air Force OÆce of Scientific Research under Grant No. AOARD-10-4053, and JSPS Grant-in-Aid for Scientific Research (C) (No. 20500147).

References 1. Castellano, C., Munoz, M.A., Pastor-Satorras, R.: Nonlinear q-voter model. Physical Review E 80, 041129 (2009) 2. Crandall, D., Cosley, D., Huttenlocner, D., Kleinberg, J., Suri, S.: Feedback e ects between similarity and sociai infiuence in online communities. In: Proceedings of KDD 2008, pp. 160–168 (2008) 3. Holme, P., Newman, M.E.J.: Nonequilibrium phase transition in the coevolution of networks and opinions. Physical Review E 74, 056108 (2006) 4. Kimura, M., Saito, K., Motoda, H.: Minimizing the spread of contamination by blocking links in a network. In: Proceedings of the 23rd AAAI Conference on Artificial Intelligence (AAAI 2008), pp. 1175–1180 (2008) 5. Kimura, M., Saito, K., Motoda, H.: Blocking links to minimize contamination spread in a social network. ACM Transactions on Knowledge Discovery from Data 3(9), 9:1–9:23 (2009) 6. Kimura, M., Saito, K., Ohara, K., Motoda, H.: Learning to predict opinion share in social networks. In: Proceedings of the 24th AAAI Conference on Artificial Intelligence (AAAI 2010), pp. 1364–1370 (2010) 7. Kleinberg, J.: Bursty and hierarchical structure in streams. In: Proceedings of the 8th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD 2002), pp. 91–101 (2002) 8. Klimt, B., Yang, Y.: The enron corpus: A new dataset for email classification research. In: Boulicaut, J.-F., Esposito, F., Giannotti, F., Pedreschi, D. (eds.) ECML 2004. LNCS (LNAI), vol. 3201, pp. 217–226. Springer, Heidelberg (2004) 9. Leskovec, J., Adamic, L.A., Huberman, B.A.: The dynamics of viral marketing. In: Proceedings of the 7th ACM Conference on Electronic Commerce (EC 2006), pp. 228–237 (2006) 10. Liggett, T.M.: Stochastic interacting systems: contact, voter, and exclusion processes. Spriger, New York (1999) 11. Palla, G., Der´enyi, I., Farkas, I., Vicsek, T.: Uncovering the overlapping community structure of complex networks in nature and society. Nature 435, 814–818 (2005) 12. Sood, V., Redner, S.: Voter model on heterogeneous graphs. Physical Review Letters 94, 178–701 (2005) 13. Swan, R., Allan, J.: Automatic generation of overview timelines. In: Proceedings of the 23rd Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR 2000), pp. 49–56 (2000) 14. Wu, F., Huberman, B.A.: How public opinion forms. In: Papadimitriou, C., Zhang, S. (eds.) WINE 2008. LNCS, vol. 5385, pp. 334–341. Springer, Heidelberg (2008) 15. Yang, H., Wu, Z., Zhou, C., Zhou, T., Wang, B.: E ects of social diversity on the emergence of global consensus in opinion dynamics. Physical Review E 80, 046108 (2009)

Dynamic Simulation of Community Crime and Crime-Reporting Behavior Michael A. Yonas, Jeffrey D. Borrebach, Jessica G. Burke, Shawn T. Brown, Katherine D. Philp, Donald S. Burke, and John J. Grefenstette Graduate School of Public Health, University of Pittsburgh, Pittsburgh, PA 15261, USA {may24,jdb108,jgburke,stb60,kdp5,donburke,gref}@pitt.edu http://www.phdl.pitt.edu

Abstract. An agent-based model was developed to explore the effectiveness of possible interventions to reduce neighborhood crime and violence. Both offenders and non-offenders (or citizens) were modeled as agents living in neighborhoods, with a set of rules controlling changes in behavior based on individual experience. Offenders may become more or less inclined to actively commit criminal offenses, depending on the behavior of the neighborhood residents and other nearby offenders, and on their arrest experience. In turn, citizens may become more or less inclined to report crimes, based on the observed prevalence of criminal activity within their neighborhood. This paper describes the basic design and dynamics of the model, and how such models might be used to investigate practical crime intervention programs. Keywords: computer simulation, agent-based model, neighborhood violence, crime prevention.

1

Introduction

While reductions in neighborhood violence have been observed, violent crime and victimization continue to be a critical issue of public safety and public health concern. The impact of violence disproportionately impacts young people living within low-income impoverished neighborhoods. For youth, violence is the second leading cause of death for youth ages 15-24 [1] and the leading cause of death for African American youth in this same age range [2]. Frequently, interventions to address community violence focus on individuals, whereas increased efforts to understand and illustrate the role of factors at all ecological levels are necessary for comprehensively addressing risk factors for neighborhood violence at the individual, family and neighborhood levels [3, 4]. Previous research has identified a variety of contextual factors such as collective efficacy, neighborhood disorganization, decreased social support, and limited resources associated with influencing neighborhood violent crime [5–8]. J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, pp. 97–104, 2011. c Springer-Verlag Berlin Heidelberg 2011 

98

M.A. Yonas et al.

Many neighborhoods affected by violence have invested heavily in developing community-level intervention responses to neighborhood violence [1, 9]. Comprehensive community mobilization interventions and such as block watch programs have been shown to be effective to address crime and violence [10, 11]. Studies assessing factors influencing citizens and victims report violence and crime have noted that a substantial number of violent crimes, approximately 50%, go unreported. Reasons for failing to report include fear of retaliation, cultural constructions related to gender, and violence in the home, work or neighborhood [12]. While some community-level approaches have shown evidence of effectiveness, these approaches are often expensive, difficult to sustain and evaluate due to lack of causal data, and challenging to replicate [11]. Modeling and simulation have the ability to evaluate potential community intervention strategies at low cost and facilitate interdisciplinary collaboration. We utilize modeling and simulation to examine and evaluate potential community crime intervention strategies. In order to inform our exploratory agent-based model, we integrate key behavioral and community factors associated with neighborhood mobilization and watch programs. This paper describes the basic design and dynamics of the model, and describes how such models might be used to investigate practical neighborhood crime intervention programs. Previous agent-based models (ABMs) have been developed to explore mitigation and dynamics of criminal activity. Epstein [13] constructed an early ABM of civil violence and rebellion with agents having levels of grievance against central authority determined by perceived-hardship and government legitimacy. The model produced a punctuated equilibrium, or periods of peace alternating with periods or rebellion. Groff [14] developed an ABM of street robbery crimes and found that inclusion of explicit geographic distributions reproduced patterns more similar to empirical patterns than other models. Furtado et al [15] utilized an ABM crime simulation including self-organizing behavior for criminals to model the swarm intelligence paradigm (i.e. most crime events in urban centers are concentrated in few places) to show that criminals’ social networks are essential for them to aggregate into organized groups. SimDrugPolicing is an ABM developed by Dray et al [16] to analyze the effects of patrol, hotspot policing and problem-oriented policing law-enforcement strategies on a street-based illegal drug market. While these previous efforts provide valuable insights into the behavior of criminals and police, the model described here contributes a novel focus on the crime-reporting behavior of citizens in a community experiencing crime.

2

Methods

An ABM was designed to study the effects of change in crime-reporting behavior as a result of exposure to incidents of crime. Both offenders and non-offenders (or citizens) were modeled as agents living in neighborhoods, with a set of rules controlling changes in behavior based on individual experience. Offenders may become more or less inclined to actively commit offenses, depending on the behavior of the neighborhood residents and other offenders, and on their experience

Dynamic Simulation of Community Crime and Crime-Reporting Behavior

99

Fig. 1. Flow chart of the activity of a citizen, for each simulation time step

with the criminal justice system. Citizens may become more or less inclined to report offenses, based on the observed prevalence of offenses within their neighborhood. The model was implemented in NetLogo 4.1. The community is represented by a two-dimensional toroidal grid; any box on the grid may be occupied by an agent. Agents move around within individuallydefined neighborhoods and may observe the behavior of other agents in their immediate surroundings. In particular, agents may witness offenses being committed and observe whether or not the offenses are punished. The behavior of an agent is controlled by a few simple rules, described in the following sections. Agents are divided into two disjoint groups: citizens, who do not commit offenses, and offenders, who may commit crimes. 2.1

Citizen Model

Citizens in the model are agents who do not commit offenses, but who may affect the prevalence of crime through their actions. The decision flow chart for a citizen is shown in Fig. 1. At each time step, the citizen moves to a random location within an individually fixed local neighborhood. Citizens may witness an offense that occurs within a given distance of the citizen’s current location. Upon witnessing a crime, the citizen experiences a witness-response that may alter future reporting behavior, and decides whether to report the crime to the police. If the citizen witnesses no crime during this time step, the citizen experiences a crime-free-response, which may also alter the citizens future reporting behavior. We represent the reporting behavior of a citizen as a two-step process: witnessing a nearby crime, and then reporting it. These steps are modeled by two probabilities for each agent i, the agents witness-probability Wi and the agents report-probability Ri . If an offense occurs within a given distance of the citizen, then the citizen witnesses the offense with probability Wi . If the citizen witnesses an offense, the agent reports the crime with probability Ri . A citizen’s

100

M.A. Yonas et al.

Fig. 2. Flow chart of the activity of an offender for each simulation time step

report-probability Ri can be affected by crime in the citizen’s neighborhood. If a crime is witnessed, the citizen experiences a witness-response: Ri = Ri ∗ (1 − α) + α, Ri = Ri ∗ (1 + α),

if if

α>0 α≤0

(1)

where α is the witness response rate. The effect of formula (1) is to raise Ri by α percent if α is positive or the decrease Ri by |α| percent if α is negative. The magnitude of α represents the tendency of a citizen to respond to an occurrence of crime by increasing the citizen’s likelihood of reporting crimes. A negative value of α reflects a situation in which a citizen becomes less likely to report crimes, due to fear of reprisal or loss of hope. Alternatively, if no crime is witnessed, the citizen experiences a crime-free-response: Ri = Ri ∗ (1 − β) + β,

if

β>0

Ri = Ri ∗ (1 + β),

if

β≤0

(2)

where β is the crime-free response rate. The magnitude of β represents the tendency of a citizen to respond to an absence of crime by increasing the citizen’s likelihood of reporting crimes. A negative value of β reflects a situation in which a citizen becomes less likely to report crimes, perhaps due to complacency. 2.2

Offender Model

Offenders in the model are agents who may commit criminal acts. The behavior of offenders during each simulation time step is shown in Fig. 2. Each offender moves around within the community seeking promising opportunities for crimes. If the current location appears attractive enough, the offender commits an offense. Upon committing an offense, the offender may be reported to the police by a witness. When an offender is reported, there is a fixed chance that the offender will be arrested.

Dynamic Simulation of Community Crime and Crime-Reporting Behavior

101

Each offender i has a probability of offending, Ci . If offender i commits a crime, the agent’s probability of offending is altered by whether or not the agent is arrested. For this model we assume that the probability of offending decreases when arrested; that is, the arrest-response is: Ci = Ci ∗ (1 − γ)

(3)

where γ ≥ 0 is the arrest response rate. Alternatively, if an offender commits a crime but is not arrested, we assume that his probability of offending increases due to his successful criminal activity; that is, the success response is: Ci = Ci ∗ (1 − δ) + δ

(4)

where δ ≥ 0 is the offender success response rate. The behavior of offenders is affected by their perception of the relative risks and opportunities in a given location. The model reflects this by associating with each location an opportunity index Op(j), defined as an estimate of the probability of committing an crime at location j and not being arrested. At each time step, each offender moves to a adjacent neighborhood location j ∗ with the maximum value of Op(j ∗ ) among all neighboring locations. Once the offender moves to location j ∗ , the decision to commit a crime is based on the comparison between the perceived opportunity at that location and the offender’s probability of offending Ci : Offender i commits a crime at current location j if Op(j) ≤ Ci . As a result, offenders tend to migrate toward those areas of the community associated with high rates of unreported criminal activity.

3

Results

The model was developed as a platform for exploring the effects of altering citizen behavior through community intervention programs. To this end, we performed an initial survey of the effects of altering citizen reporting-behavior in the model by systematically altering the witness response rate α and the crime-free response rate β. Witness response α might be increased through interventions such as decreasing tolerance for neighborhood crime through the removal of graffiti, and making it easier to report crimes by providing neighborhood phone-trees (allowing citizens to call a neighbor who reliably contacts the police). The crime-free response β might be increased by programs targeted to raise expectations of a crime-free neighborhood and decrease citizen complacency, such as encouraging citizens to spend more time outdoors communicating with neighbors. A series of simulations was run, varying α from -0.2 to 0.2 and varying β from -0.02 to 0.02. (Initial studies showed that the results of the model do not vary significantly outside these parameter ranges). The values of fixed parameters are shown in Table 1. For each scenario, 10 replications of the model with identical initial conditions were performed. Figure 3 shows mean results for (a) total crime, (b) proportion

102

M.A. Yonas et al. Table 1. Fixed Parameters for Crime ABM Simulation Parameter

Value

Number of Citizens Number of Offenders Number of Locations Witness Probability (Wi ) Initial Report Probability (Ri ) Initial Probability of Offending (Ci ) Probability of Arrest when Reported Arrest Response Rate (γ) Offender Success Response Rate (δ) Jail Term Length of Simulation

900 100 10,000 U(0,0.5) U(0,1) U(0,1) 0.5 0.1 0.5 U(1,30) days 365 days

U(a,b): uniform distribution between a and b

of the community that has been crime-free for 30 days, and (c) number of active citizens (defined as citizens whose probability of reporting a crime is greater than 50%). The results show some expected patterns. For example, total crime is minimized in the model when large positive values of α and β are observed, which represent citizens becoming more active reporters regardless of changes in crime prevalence in their neighborhood. Perhaps a more surprising result occurs when α = 0.2 and β = −0.02 (lowest corner in the plots). These values correspond to scenarios in which citizens become more likely to report when crime occurs, but

Fig. 3. Summary results over variations in values for α and β. (a) Total crime. (b) Final proportion of crime-free area, over variations in values for α and β. (c) Final number if active citizens.

Dynamic Simulation of Community Crime and Crime-Reporting Behavior

103

less likely to report when the neighborhood becomes crime-free. This scenario yields an expected low total crime (Fig 3a), relatively high crime-free area (Fig 3b) and interestingly requires surprisingly few citizens who are likely to report crime (Fig 3c). Assuming that the overall cost of an intervention program may be correlated with the number of citizens whose reporting behavior needs to change, this scenario provides support for community watch crime prevention programs.

4

Discussion

The current model provides useful insights into which interventions at the neighborhood level might be most effective in reducing the presence of crime in a community. Future studies will elaborate the model to reflect more detailed behavior for both citizens and offenders (e.g., retaliatory behavior). For example, in developing this formative model, the arrest probability was fixed at 50%. In future models reflecting real-world settings, α, β, γ, and δ are likely to be functions of the probability of citizen action and arrest. Future models will also include explicit elements of the law enforcement system (e.g., increasing police presence in neighborhoods experiencing high crime rates). Of course, it is recognized that computational models cannot reflect the complexity of actual communities and of individual responses to crimes within a community. Nonetheless, such models provide an additional set of tools for generating inclusive dialogue to inform the development of intervention policies which equitably integrate interdisciplinary expertise of neighborhood residents, law enforcement, policy makers and researchers. As more realistic computational models are developed, it may be possible to virtually evaluate a broad range of possible interventions with unique multi-level characteristics, and potentially identify those that may be likely to be successful in preventing community crime. Acknowledgments. Supported in part by the National Institute of General Medical Sciences MIDAS grant 1U54GM088491-01.

References 1. Center for Disease Control and Prevention, National Center for Injury Prevention and Control. Web-based Injury Statistics Query and Reporting System (WISQARS), www.cdc.gov/injury/wisqars/ (accessed 2010 June 14) 2. Center for Disease Control and Prevention, National Center for Injury Prevention and Control. Web-based Injury Statistics Query and Reporting System (WISQARS), Youth Violence Data Sheet, www.cdc.gov/violenceprevention/ pdf/YV-DataSheet-a.pdf (accessed 2010 June 14) 3. Kellerman, A.L., Fuqua-Whitley, D.S., Rivara, F.P., Mercy, J.: Preventing Youth Violence: What Works? Annual Review of Public Health 19, 271–292 (1998) 4. National Youth Violence Prevention Resource Center (NYVPRC), www.safeyouth.gov/Resources/Prevention/Pages/PreventionStrategies. aspx

104

M.A. Yonas et al.

5. Anderson, E.: Code of the Street: Decency, Violence and the Moral Life of the Inner City. Norton, New York (1999) 6. Earls, F.J.: Violence and Todays Youth: Critical Health Issues for Children and Youth. Future of Children 4(3), 4–23 (1994) 7. Sampson, R.J., Raudenbush, S.W., Earls, F.: Neighborhoods and Violent Crime: A Multilevel Study of Collective Efficacy. Science 277, 918–924 (1997) 8. Taylor, R.B., Gottfredson, S.D., Brower, S.: Understanding Block Crime and Fear. Journal of Research in Crime and Delinquency 21, 303–331 (1984) 9. Wilson, J.M., Chermak, S., McGarrell, E.F.: Community-Based Violence Prevention: An Assessment of Pittsburgh’s One Vision One Life Program. RAND Corp., Santa Monica (2010) 10. Bibb, M.: Gang Related Services of Mobilization for Youth. In: Klein, M.W. (ed.) Juvenile Gangs in Context: Theory, Research, and Action. Prentice-Hall, Englewood Cliffs (1967) 11. Bennett, T.H., Holloway, K.R., Farrington, D.P.: Effectiveness of Neighbourhood Watch in Reducing Crime. National Council on Crime Prevention, Stockholm (2008) 12. Reiss, A.J., Roth, J.A.: Measuring Violent Crime and Their Consequences. In: Understanding and Preventing Violence. National Research Council, pp. 404–429. National Academy Press, Washington DC (1993) 13. Epstein, J.: Modeling Civil Violence: An Agent-based Computational Approach. Proceedings of the National Academy of Sciences of the United States of America 99 (suppl. 3), 7243–7250 (2002) 14. Groff, E.: Characterizing the Spatio-temporal Aspects of Routine Activities and the Geographical Distribution of Street Robbery. In: Liu, L., Eck, J. (eds.) Artificial Crime Analysis Systems, pp. 226–251. IGI Global, Hershey (2008) 15. Furtado, V., Melo, A., Coelho, A.L.V., Menezes, R., Belchio, M.: Simulating Crime against Properties using Swarm Intelligence and Social Networks. In: Liu, L., Eck, J. (eds.) Artificial Crime Analysis Systems, pp. 300–318. IGI Global, Hershey (2008) 16. Dray, A., Mazerolle, L., Perez, P., Ritter, A.: Drug Law Enforcement in an AgentBased Model: Simulating the Disruption. In: Liu, L., Eck, J. (eds.) Artificial Crime Analysis Systems, pp. 352–371. IGI Global, Hershey (2008)

Model Docking Using Knowledge-Level Analysis Ethan Trewhitt, Elizabeth Whitaker, Erica Briscoe, and Lora Weiss Georgia Tech Research Institute, 250 14th Street NW, Atlanta, GA 30332-0832 {ethan.trewhitt,elizabeth.whitaker,erica.briscoe, lora.weiss}@gtri.gatech.edu

Abstract. This paper presents an initial approach for exploring the docking of social models at the knowledge level. We have prototyped a simple blackboard environment allowing for model docking experimentation. There are research challenges in identifying which models are appropriate to dock and the concepts that they should exchange to build a richer multi-scale view of the world. Our early approach includes docking of societal system dynamics models with individual and organizational behaviors represented in agent-based models. Casebased models allow exploration of historical knowledge by other models. Our research presents initial efforts to attain opportunistic, asynchronous interactions among multi-scale models through investigation and experimentation of knowledge-level model docking. A docked system can supply a multi-scale modeling capability to support a user’s what-if analysis through combinations of case-based modeling, system dynamics approaches and agent-based models working together. An example is provided for the domain of terrorist recruiting. Keywords: model docking, knowledge level modeling, federated models, socio-cultural modeling.

1 Introduction Modeling complex, social behaviors requires representations provided by different modeling approaches and across multiple scales [1]. An analyst may need information about individual, organizational, and societal behaviors in a system model that explores the interactions of variables at each of these scales of behavior. These federated models need to interoperate or become docked, i.e., communicate, collaborate, and share asynchronous information, to represent the system being studied. This paper focuses on developing a knowledge-level or semantic approach to federated model interoperability. There are many approaches for enabling the interaction of models at the data level, such as the High Level Architecture (HLA) for distributed simulations or MATREX1 for battlespace modeling. This paper presents a research approach to enabling the interaction of multi-scale models at the knowledge-level. The specific advances that are made in this paper include presenting the types of knowledge (as opposed to specific data structures) that need to be exposed across models of different scales to enable seamless interoperability. 1

Modeling Architecture for Technology Research & Experimentation.

J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, pp. 105–112, 2011. © Springer-Verlag Berlin Heidelberg 2011

106

E. Trewhitt et al.

There are many research challenges in identifying which models are appropriate to dock and the relevant concepts that they should exchange to build a richer multi-scale view of the world being modeled. Our early experiments [2] include docking of societal system dynamics models [3] with individual and organizational behaviors represented in agent-based models [4]. More recently, we have expanded our docking approach to include the docking of case-based models [5], which allow exploration of historical knowledge by other models. Macro-scale models of a society, e.g., system dynamics-based models, rarely track the movements or opinions of individuals, and instead work with trends of the overall population. System dynamics models are useful for monitoring the effects of policies or actions on broad demographic groups. However, the decisions of people within the population are nuanced and are often dependent on those individuals’ personal experiences, and individual traits are best simulated by a micro-scale model, such as an agent-based model. Historical experiences are best captured by a case-based reasoning system and can be used to inform the building of an agent that behaves rationally. Each of these model types has advantages and disadvantages. Rather than try to capture an entire system’s behavior with a single model type, docking allows each model to represent the scale and type of information for which it is best suited. The research presented in this paper consists of our initial efforts to attain this goal through the investigation and experimentation of knowledge-level model docking.

2 Model Docking Framework 2.1 Existing Model Docking Frameworks Existing model docking frameworks, such as the US Defense Advanced Research Projects Agency’s COMPOEX2 [6], offer effective tools for implementing the communication-layer and timing details necessary to create federated models. The Army’s OneSAF (One Semi-Automated Forces) modeling environment combines tactical operations up to the battalion and brigade levels, with variable levels of resolution and supports composability within a single simulation environment. The Stabilization and Reconstruction Operations Model (SROM) supports analyzing interdependencies of PMESII3 variables, but it is still predominantly a macro-scale model. Although there has been substantial work to enable multi-scale analyses and communication through a shared vocabulary or ontology, there is still much work to be done in understanding how to enable interactions of multiple models developed by different researchers, where the interactions occur at the knowledge-level and across different scales. The research presented in this paper aims to improve understanding these knowledge-level interactions. 2.2 The Knowledge Level The Knowledge Level is a concept developed by Newell [7] to describe a level of representation above the symbolic level. At the knowledge-level, the research focuses 2 3

Conflict Modeling, Planning, and Outcomes Experimentation. Political, Military, Economic, Social, Infrastructure, and Information.

Model Docking Using Knowledge-Level Analysis

107

on the study of the interaction of model content and semantics rather than the data structures, algorithms, or implementation used to enable the interaction. Our research focuses on exposing information between models at the knowledge-level. Thus, a mechanism for enabling knowledge-level experimentation for model docking is necessary. We use a blackboard approach to support the loose integration of model components through a shared knowledge area and enable the opportunistic use of shared knowledge by federated components. This allows for an easily reconfigurable experimentation environment enabling the research focus to remain on understanding the content to be exposed at the knowledge level. 2.3 Communicating across a Blackboard The Blackboard [8, 9] environment uses a shared working memory area, called the blackboard, for collaborative problem solving amongst a set of autonomous knowledge sources that write information to and read information from the blackboard. Blackboard frameworks were developed to address complex and ill-defined problems for which existing techniques were inadequate. The knowledge sources are typically executable models or model components. They can post partial solutions, e.g., information that they have learned or generated, that may be of use to other knowledge sources. This information may include goals (or subgoals) and partial or complete problem solutions. The knowledge sources post requests for services or information from other knowledge sources and may choose to respond to requests. In this way, the blackboard approach supports asynchronous, opportunistic problem-solving by a “team” of collaborating knowledge sources. Exploring and prototyping model docking using a blackboard allows models of different types and scales to post summaries or interpretations of scenarios or situations that may be useful to other models. The asynchronicity of the blackboard enables models to take information when they need it and based on their time-scales. The blackboard approach is one that inherently enforces very little structure upon the models that make use of the shared memory area. This makes it an ideal method of communicating and sharing knowledge in an experimental system where the concepts being exchanged and the problem solving collaborations are being reconfigured in order to explore model docking issues. This paper shows how a blackboard docking approach can be used for linking agent-based, system dynamics, and case-based models. This approach helps address the information gap caused by separate models and enables systematic reasoning of individual, social, and cultural behaviors. It also enables interactions between dynamic models resulting in increased capabilities for analysis and interpretation of behaviors, and it facilitates what-if analyses and analytical exploration associated with understanding multi-scale behavior. 2.4 Model Considerations When developing a federated model, several questions must be kept in mind so as to ensure the results are relevant to the analysis. These questions helped support the development of the federated approach presented in this paper:

108

E. Trewhitt et al.

• What is the use case for the overall federated model? • What are the individual knowledge-level elements that require modeling? • What type of model is best for representing each system component? (Some models may represent multiple components.) • What knowledge does each model require from other models? • What knowledge can each model contribute to other models? • What connecting concepts exist between each pair of models?

3 Docking Concept and Example Execution Fig. 1 presents a graphic of three model types docked through blackboard. System Dynamics Model

Translation Module

Aggregate Agent Decisions

Societal Trends, Policies, Events

Agent-Based Model

Case-Based Reasoner

Agents

Case Library

Translation Module

Societal Trends, Group Influences

Agent Decisions, Opinions

Translation Module

Requests for Case Retrieval

Adapted Case Results

Blackboard

Analyst User Interface

Development Observer

Fig. 1. Example of three model types docked with a blackboard

3.1 Identifying the Connecting Concepts In building a complex multi-scale docked model, the individual models should be chosen so that concepts and vocabularies provide different views of the same system. The connecting concepts must be identified. These are the concepts represented in each model that could be supplied to other models docked to enable a richer representation of the system. An example is presented to illustrate this approach. 3.2 Example Use Case To illustrate the approach, we present an example use case built around a terrorist recruiting scenario that includes elements at three different scales of analysis: micro, meso, and macro. The federated system is meant to answer questions about terrorist recruitment patterns. Four independent executable models are included, each of which performs a specialized task:

Model Docking Using Knowledge-Level Analysis

109

─ Agent-Based Models. Two agent-based models are included, one with agents simulating individuals and one with agents representing organizations. The individual agents make a recruitment decision based on their own personal traits and experiences. Group agents (representing terrorist groups, religious organizations, and community groups) accept societal influences and exert influence over their members and other groups. ─ Case-Based Model. A case-based model is included that provides a set of past cases of individual recruiting. Each case consists of a single individual targeted for recruitment, including the properties of the individual, the group doing the recruiting, and the context. The solution consists of the individual's decision whether to join the group. ─ System Dynamics Model. A system dynamics model is included that provides a representation of societal interactions. This model represents the context in which the individuals and organizations make their decisions, including variables such as economic and political stability, government policy constraints, and leadership popularity. Individually, these models are capable of modeling only a piece of the overall system, but in a federation, they provide a much broader environment for analysis. Figure 2 presents an informal, example schedule of communications between models, as well as the expected knowledge to be passed from one model to another. Each model’s output represents the most relevant information that it has produced, rather than being customized for use by a specific model. Because model communications are loosely defined, modifications to the simulation require less modification of the models themselves. In each model interaction, the sending model’s output is written to the blackboard, while the receiving model reads the data from the blackboard. It is the obligation of the sending model to create knowledge that is as broadly useful as possible, while it is the obligation of the receiving model to use that knowledge for its own needs. To understand how the system cycles and executes, we walk through the process in Fig. 2. Here, an analyst sets up execution parameters for baseline values that are relevant across all scales of knowledge. The end goal is for the analyst to use the system to perform what-if experiments by altering input parameters and monitoring changes in the outcome of the overall simulation. Once the inputs for a particular simulation have been entered, the system begins execution. The execution occurs as follows. Input parameters are passed to the appropriate sub models for initiation. The system dynamics model begins by simulating macroscale (society-level) variables to support agent creation. Agents are used in two ways. They are created to represent a hypothetical population (meso-scale behavior) using demographics determined by the system dynamics model, and they are created to represent individual (micro-scale behavior) that affect individual decisions. The casebased reasoning model retrieves and adapts cases in which past stories are similar to the current individuals and context. For cases that do not satisfy a certain similarity metric from those available in the case library, the agent-based models prepare to simulate the responses of those individuals using an internal influence model. The agent-based knowledge for the historical cases and simulated individuals are aggregated and used as numerical values by the system dynamics model. The system dynamics model outputs societal decisions to the agents.

110

E. Trewhitt et al.

UI Input

UI Output

Context

User (analyst)

SysDyn (macro)

Execute

Results (Iterations)

Adapt

Macro numbers

Multi-Case Lookup

Simulate

Micro and meso behaviors

Multiple Profiles

Setup

Context

CBR

Multi-person Sim. Params.

ABM (meso)

Simulate

Micro behaviors

Setup

Context

Setup

Multi-profile Request

ABM (micro)

Execute

Fig. 2. Sample timing diagram of example model interactions Table 1. Example connecting concepts Model Agent model, individual

Case-based model

System dynamics model

Inputs ─ Societal trends ─ Conservatism of government ─ Opinions of society: gov’t, morality, etc. ─ Events occurring in environment ─ Economic conditions ─ Targeted recruitment efforts ─ Influence of leaders, friends ─ Request for case retrieval ─ An individual’s situation ─ Details of recruitment efforts ─ Question: “Is this person likely to be recruited?” ─ Aggregate individual behaviors ─ Decisions to be recruited ─ Opinions of government, coalition actions ─ Aggregate details about environment, events

Outputs ─ Opinions about terrorist group ─ Decision whether to join terrorist group ─ Interaction history

─ Solution: yes or no ─ Explanation of solution? ─ Probabilities ─ Trends about societal decisions, opinions ─ Trends about recruitment efforts ─ Cultural variables

The system cycles amongst all the models, simulating individual attitudes, group behaviors, and societal responses, until the simulation reaches the terminating condition specified by the analyst. This configuration allows the docked models to autonomously control their execution and timing, referencing the overall modeling timeline and to allow meaningful information exchange across the models with differing time scales. This is known as opportunistic model execution. An important element of this approach is visibility. Since a goal is to enable experimentation with model docking approaches, it is important to be able to view

Model Docking Using Knowledge-Level Analysis

111

the explicit knowledge exchanges among models to understand and evaluate the connecting concepts. Since each model’s output is written to the blackboard for other models’ use, a module, called the development observer, is incorporated to enable monitoring of interactions between models and to display the interactions to the model developer. This information can help the developer improve the design of such interactions. Table 1 shows some of the example connecting concepts. 3.3 Information to Expose One of the benefits of docking independent models is that factors that influence behavior can exist at their ‘natural’ scale, i.e., the scale where they ordinarily occur. There is enough consistency to identify common, critical behavioral influences at their natural levels of representation. Typical macro-scale information includes demographics, culture, economics, and political structure. Also significant at this level is the representation of events affecting a large number of people or occurring over a long period a time. Representation at the meso-scale may include groups and organizations which may hold organizational beliefs and demonstrate behaviors. The structure of groups is also an important characteristic which may be used to determine how micro-scale agents communicate and are influenced by meso-scale variables. At the micro-scale, agents may represent individuals and typically demonstrate behaviors that are the result of the individuals’ beliefs and attributes. Behaviors are especially relevant to other scales of analysis, i.e., many individual decisions may be aggregated to create a higher-scale attribute. Micro-scale aggregation of beliefs can result in a different representation of beliefs than when perceived from the meso-scale representation of group beliefs. A loosely structured classification of the types of information most commonly shared between scales should facilitate an analytical approach docking models that represent different behavioral aspects and scales of the same general problem. A recent paper [10] provides a more comprehensive description of the information types to expose at each scale.

4 Conclusions This paper presents an initial approach for exploring the docking of social models at the knowledge level. We have prototyped a simple blackboard environment that allows for experimentation with docking models and iterative improvement of the approach. There are many research challenges in identifying which models are appropriate to dock and the relevant concepts that they should exchange to in order to build a richer multi-scale view of the world being modeled. Our early approach includes docking of societal system dynamics models with individual and organizational behaviors represented in agent-based models and case-based models, which allow exploration of historical knowledge by other models. The research presented in this paper represents initial efforts to attain opportunistic and asynchronous interactions among multi-scale models through the investigation and experimentation of knowledge-level model docking.

112

E. Trewhitt et al.

References 1. Zacharias, G., MacMillan, J., Van Hemel, S. (eds.): Behavioral Modeling and Simulation: from Individuals to Societies. National Research Council, Washington, D.C. (2008) 2. Weiss, L., Whitaker, E., Briscoe, E., Trewhitt, E.: Cultural and Behavioral Model Docking Research. In: Human Social Cultural and Behavioral Modeling Focus 2010 Conference, Chantilly, VA (2009), http://www.gtri.gatech.edu/files/media/ Weiss_Abstract_HSCB_Conference_2010_public.pdf 3. Forrester, J.: Counterintuitive Behavior of Social Systems. Technology Review 73(3), 52– 68 (1971) 4. Macal, C., North, M.: Tutorial on Agent-Based Modeling and Simulation. In: Proceedings of the 2005 Winter Simulation Conference, Argonne (2005) 5. Kolodner, J.: Case-Based Reasoning. Morgan Kaufmann, San Mateo (1993) 6. Defense Advanced Research Projects Agency COMPOEX, http://www.darpa.mil/ipto/programs/compoex/compoex.asp 7. Newell, A.: The Knowledge Level. AI Magazine 2(2), 1–33 (1981) 8. Engelmore, R.S., Morgan, A. (eds.): Blackboard Systems. Addison-Wesley, Reading (1988) 9. Whitaker, E., Bonnell, R.: A Blackboard Model for Adaptive and Self-Improving Intelligent Tutoring Systems. Journal of Artificial Intelligence in Education (1992) 10. Briscoe, E., Trewhitt, E., Whitaker, E., Weiss, L.: Multi-Scale Interactions in Behavior Modeling, Submitted to: Social Computing, Behavioral-Cultural Modeling and Prediction 2011, March 29-31, Sundance, UT (2011)

Using Web-Based Knowledge Extraction Techniques to Support Cultural Modeling Paul R. Smart1 , Winston R. Sieck2 , and Nigel R. Shadbolt1 1

2

School of Electronics and Computer Science, University of Southampton, Southampton, SO17 1BJ, UK {ps02v,nrs}@ecs.soton.ac.uk Applied Research Associates, Inc., 1750 Commerce Center Blvd, N., Fairborn, OH, 45324, USA [email protected]

Abstract. The World Wide Web is a potentially valuable source of information about the cognitive characteristics of cultural groups. However, attempts to use the Web in the context of cultural modeling activities are hampered by the large-scale nature of the Web and the current dominance of natural language formats. In this paper, we outline an approach to support the exploitation of the Web for cultural modeling activities. The approach begins with the development of qualitative cultural models (which describe the beliefs, concepts and values of cultural groups), and these models are subsequently used to develop an ontology-based information extraction capability. Our approach represents an attempt to combine conventional approaches to information extraction with epidemiological perspectives of culture and network-based approaches to cultural analysis. The approach can be used, we suggest, to support the development of models providing a better understanding of the cognitive characteristics of particular cultural groups. Keywords: cultural modeling, ontology-based information extraction, culture, cognition, knowledge extraction, world wide web.

1

Introduction

The World Wide Web (WWW) is a valuable source of culture-relevant information, and it is therefore an important resource for those interested in developing cultural models. The exploitation of the WWW in the context of cultural modeling is, however, hampered both by the large-scale nature of the Web (which makes relevant information difficult to locate) and the current dominance of natural language formats (which complicates the use of automated approaches to information processing). In this paper, we describe an approach to support the use of the Web in cultural modeling activities. The approach is based on the development of ontology-based information extraction capabilities, and it combines the use of Semantic Web technologies and natural language processing (NLP) techniques with an epidemiological perspective of culture [1] and the use J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, pp. 113–120, 2011. c Springer-Verlag Berlin Heidelberg 2011 

114

P.R. Smart, W.R. Sieck, and N.R. Shadbolt

of network-based approaches to cultural analysis [2]. Technological support for the approach is currently being developed in the context of the IEXTREME1 project, which is funded by the U.S. Office of Naval Research. The structure of the paper is as follows. In Section 2, we outline what is meant by the term ‘culture’, and we describe an approach to cultural modeling that is based on the development of models representing the ideas associated with particular cultural groups. In Section 3, we describe our approach to the development of Web-based knowledge extraction capabilities to support cultural model development. This approach combines conventional approaches to information extraction with semantically-enriched representations of cultural models, and it seeks to provide a culture-oriented ontology-based information extraction (OBIE) capability for the WWW.

2

Cultural Models and Cultural Network Analysis

Before addressing the use of the Web to study culture, we first need to define what is meant by the term ‘culture’. As is to be expected in any highly interdisciplinary field, there are a variety of conceptions of culture. Our conception is distinctly cognitive in nature, and it is based on an epidemiological perspective [1]. A fundamental assumption of this perspective is that shared developmental experiences lead to important similarities in the mental representations (e.g. values and causal knowledge) that are distributed among members of a population. Culturally widespread ideas ground the distribution of behavioral norms, discussions, interpretations, and affective reactions, and researchers working within the epidemiological perspective thus seek to describe and explain the prevalence and spread of ideas within specific populations. Working from this perspective, we previously developed a technique called cultural network analysis (CNA), which is a method for describing the ‘ideas’ that are shared by members of cultural groups [2]. CNA discriminates between three kinds of ideas, namely, concepts, values, and causal beliefs, which together constitute the contents of what are called ‘cultural models’. These cultural models typically rely on the use of belief network diagrams to show how the set of relevant ideas relate to one another (see Fig. 1 for an example). In general, we can distinguish two types of cultural models: qualitative and quantitative cultural models (see [2]). Qualitative cultural models present the ideas associated with a particular group, whereas quantitative models add information about the prevalence of those ideas in the target population. In addition to seeing the approach described in this paper as a means to validate and refine qualitative cultural models, it is also possible to see the approach as enabling a cultural analyst to estimate the relative frequency of ideas in a target population and thus develop quantitative extensions of qualitative cultural models (see Section 3.6).

1

See http://www.ecs.soton.ac.uk/research/projects/746

Web-Based Knowledge Extraction and Cultural Modeling

3

115

Web-Based Knowledge Extraction for Cultural Model Development

In this section, we describe an approach to cultural model development that combines CNA with state-of-the-art approaches to knowledge representation and Web-based information extraction. The aim is to better enable cultural model developers to exploit the WWW as a source of culture-relevant information. The approach we describe is based on a decade of research into OBIE systems (see [3] for a review), and it combines conventional approaches to information extraction with an ontology that provides background knowledge about the kinds of entities and relationships that are deemed important in a cultural modeling context. The first step in the process is to develop an initial qualitative cultural model using a limited set of knowledge sources (see [2] for more details on this step). The second step involves the development of a cultural ontology using the qualitative cultural model as a reference point. This ontology is represented using the Ontology Web Language (OWL), which has emerged as a de facto standard for formal knowledge representation on the WWW. The third step is to manually annotate sample texts using the cultural ontology in order to provide a training corpus for rule learning. Rule learning, in the current context, is mediated by the (LP)2 algorithm, which is a supervised algorithm that has been used to develop a variety of adaptive information extraction and semantic annotation capabilities [4,5]. Following the development of information extraction rules, the rules are then applied to Web resources in the fourth step in order to identify instances of the entities defined in the initial qualitative cultural model. Step five consists in the identification and extraction of causal relations. The extraction of causal relationships is a difficult challenge because techniques for information extraction have tended to focus on the extraction of particular entities in a text, rather than the relationships between those entities. We attempt to extract causal relationships using an approach that combines the use of background knowledge in the form of a domain ontology with the general purpose lexical database, WordNet [6]. Finally, in step six, the extracted cultural knowledge is integrated, stored, and used to estimate the relative frequencies of the various ideas presented in the initial qualitative cultural model. We briefly describe each of these steps in subsequent sections. 3.1

Step 1: Develop Qualitative Cultural Model

The technique used to develop qualitative cultural models has been described in previous work [2], and we will not reiterate the details of the technique here. Fig. 1 illustrates a simplified qualitative cultural model that represents an extremist Sunni Muslim’s beliefs about current socio-political relationships between Islam and the West. The set of ideas represented in Fig. 1 were extracted from articles that describe jihadist narratives, and they are presented here for illustrative purposes. The cultural model illustrates concepts shared by the group, as well as their common knowledge of the causal relationships between those concepts. This shared knowledge influences expectations about how socio-political

116

P.R. Smart, W.R. Sieck, and N.R. Shadbolt

Fig. 1. Sunni extremist cultural model of jihad (simplified)

relationships will unfold, and it provides a basis for the selection of particular actions and decision outcomes; for example, the decision to support jihad. Fig. 1 shows the three different kinds of ideas that are the targets of a cultureoriented knowledge extraction system. These ideas include simple concepts such as “Western ideology” and “Muslim Honor”, each represented as closed shapes in Fig. 1. They also include causal beliefs; for example, the idea that Western ideology (e.g. secularism, nationalism) is inhibiting the formation of a unified Islamic caliphate and the idea that the West promotes this ideology because it is engaged in a covert war against Islam. These causal beliefs are represented as arrows in the figure, with the +/- symbols indicating the polarity of the causal relationship. Finally, Fig. 1 portrays values using specific shapes, with circles indicating “positive” outcomes and hexagons indicating “negative” outcomes. Developing an Islamic caliphate is thus a good thing according to the cultural model. Maintaining (and enhancing) Muslim honor is likewise valued. According to the model, jihad is viewed positively and should be supported by the model’s adherents due to the perceived anticipated consequences for Muslims. Most directly, support for jihad decreases the chances that the West will continue its war against Islam, and it enhances collective Muslim honor. 3.2

Step 2: Develop Cultural Ontology

Once an initial qualitative cultural model has been developed, the next step in the process is to develop an ontology that represents the contents of the model. The main reason this step is undertaken is that it enables the cultural model to be used to support information extraction. Over recent years, a rich literature has emerged concerning the use of ontologies in information extraction, and a number of important tools have been developed to support OBIE [3]. By converting the cultural model into an ontology using standard knowledge representation languages, such as OWL, we are able to capitalize on the availability of these pre-existing tools and techniques, and we are also able to compare the success of our approach with other OBIE approaches.

Web-Based Knowledge Extraction and Cultural Modeling

117

The ontologies developed to represent the contents of cultural models are based around the notion of ideas as being divided into concepts, causal beliefs and values. These three types of ideas constitute the top-level constructs of the ontology, and subtypes of these constructs are created to represent the various elements of the cultural model. For example, if we consider the notion of “Jihad support”, as depicted in Fig. 1, then we can see that this is a type of concept, and it is regarded as a positive thing, at least from the perspective of the target group. Within the ontology developed to support this cultural model we have the concept of ‘support-for-jihad’, which is represented as a type of ‘jihadrelated-action’, which is in turn represented as a type of ‘action’, which is in turn represented as a type of ‘concept’. Given the focus of the cultural model in representing causal beliefs, the notions of actions, events and the causallysignificant linkages between these types of concept are often the most important elements of the cultural ontology. 3.3

Step 3: Develop OBIE Capability

In this step of the process, the aim is to create rules that automatically detect instances of the concepts, beliefs and values contained in the cultural model. There are clearly a number of ways in which this might be accomplished, especially once one considers the rich array of information extraction techniques and technologies that are currently available [7]; however, we prefer an approach that delivers symbolic extraction rules (i.e. rules that are defined over the linguistic features and lexical elements of the source texts) because the knowledge contained in the rules can be easily edited by subject matter experts. In addition, it is easier to provide explanation-based facilities for rule-based symbolic systems than it is for systems based on statistical techniques. The approach to rule creation that we have adopted in the context of the IEXTREME project is based on the use of the (LP)2 learning algorithm, which has been used to create a number of semantic annotation systems [4,5]. The basic approach is to manually annotate a limited number of source texts using the cultural ontology that was created in the previous step. These annotated texts are then used as the training corpus for rule induction (see [8] for more details). During rule induction, the (LP)2 algorithm generalizes from an initial rule that is created from a user-defined example by using generic shallow knowledge about natural language. This knowledge is provided by a variety of NLP resources, such as a morphological analyzer, a part-of-speech (POS) tagger and a gazetteer. The rules that result from the learning process thus incorporate a variety of lexical and linguistic features. Previous research has suggested that rules can be defined over a large number of features. For example, Bontcheva et al [9] used a variety of NLP tools to generate 94 features over which information extraction rules could be defined. Of course, not all of these features are likely to be of equal importance in creating information extraction systems, and further empirical studies are required to assess their relative value in the domain of cultural modeling.

118

3.4

P.R. Smart, W.R. Sieck, and N.R. Shadbolt

Step 4: Extract Concepts

Once extraction rules have been created, they can be applied to potential knowledge sources in order to detect occurrences of the various ideas expressed in the cultural model. Because most of the user-defined annotations will be based on the nodal elements of the cultural model networks, such as those seen in Fig. 1, this step is particularly useful for detecting mentions of specific concepts in source texts. In the case of Sunni extremist cultural models this could, for example, include mentions of jihad-related concepts, for example “Jihad is a means to expel the Western occupiers”, as well as references to aspects of Western ideology. In general, information extraction systems based on the machine learning technique described above (i.e. the (LP)2 algorithm) have proved highly effective in identifying instances of the terms defined in an ontology, so we can expect reasonable extraction performance for this step of the process. 3.5

Step 5: Extract Causal Relations

There have been a number of attempts to extract relational information in a Webbased context (see [7]). The use of ontologies in such systems plays an important role because they provide background knowledge about the possible semantic relationships that are likely to exist between the various entities identified in previous processing steps. Thus, if a system first subjects a textual resource to entitybased semantic annotation, then it is able to use the ontology to form expectations about the kind of relationships that might be apparent in particular text fragments. When this background knowledge is combined with lexical and linguistic information, a relation extraction system is often able to identify relationships that would be impossible to detect using a text-only analysis. The approach to relation extraction that we have adopted in the case of the IEXTREME project is based on a technique that was previously developed to support information extraction in the domain of artists and artistic works [10]. The approach builds on the outcome of the previous step, which is concerned with the detection of concepts in the source texts. Importantly, once these concept annotations are in place, the relation extraction subsystem is provided with a much richer analytic substrate than would otherwise have been the case. In fact, it is only once such annotations are in place that the real value of the ontology (for the detection and extraction of relationships) can be appreciated. In particular, the ontology provides background knowledge that drives the formation of expectations about the kinds of relationships that could appear between concepts, and once such expectations have been established, they can be supported or undermined by subsequent lexical analysis of the sentence in which the concepts occur. Obviously, the nature of the natural language processing that is performed on the sentence is key to this relation-based annotation capability. It is not sufficient for a system to simply form an expectation about the kind of relationships that might occur between identified entities in the text; the system also needs to ascertain whether the linguistic context of the sentence supports the

Web-Based Knowledge Extraction and Cultural Modeling

119

assertion of a particular relationship. The decision regarding which relationship (if any) to assert in a particular sentential context is based on a strategy similar to that used in previous research [10]. Essentially, each relationship in the ontology is associated with a ‘synset’ (a set of synonyms) in the general-purpose lexical database WordNet [6]. When the relation extraction system executes, it attempts to match the words in a sentence against the WordNet-based linguistic grounding provided for each expected relationship. In addition to representing information about synonyms, the WordNet database also represents hypernymy (superordinate) and hyponymy (subordinate) relationships. These can be used to support the matching process by avoiding problems due to transliteration. 3.6

Step 6: Exploit Knowledge Extraction Capability

The knowledge extraction capability outlined in the previous steps provides support for the refinement, extension and validation of the knowledge contained in cultural models. The ability to detect instances of the ideas expressed in cultural models across a range of Web resources (including blogs, organizational websites, discussion threads and so on) provides a means by which new knowledge sources can be discovered and made available for a variety of further model development and refinement activities. The use of OBIE technology therefore provides a means by which the latent potential of the Web to serve as a source of culture-relevant knowledge and information can be exploited in the context of qualitative cultural modeling initiatives. Aside from the development of better qualitative cultural models, the use of knowledge extraction techniques can also support the development of quantitative cultural models. As discussed above, quantitative cultural models extend qualitative cultural models by including information about the relative frequencies of particular ideas within the population to which the cultural model applies [2]. By harnessing the power of OBIE methods, the current approach provides a means by which ideas (most notably concepts and causal beliefs) can be detected across many hundreds, if not thousands, of Web resources. This provides an estimate of the prevalence of particular ideas in the target population of interest, and it provides a means by which the Web can be used to support the development of quantitative cultural models.

4

Conclusion

This paper has described an approach to harnessing the latent potential of the Web to support cultural modeling efforts. The approach is based on the development of culture-oriented knowledge extraction capabilities and the use of techniques that support a cognitive characterization of specific cultural groups. Systems developed to support the approach may be seen as an important element of iterative cultural modeling efforts, especially ones in which an initial qualitative cultural model drives the acquisition of information from a large number of heterogeneous Web-based resources.

120

P.R. Smart, W.R. Sieck, and N.R. Shadbolt

Acknowledgments. This research was supported by Contract N00014-10-C0078 from the U.S. Office of Naval Research.

References 1. Sperber, D.: Explaining Culture: A Naturalistic Approach. Blackwell, Malden (1996) 2. Sieck, W.R., Rasmussen, L., Smart, P.R.: Cultural Network Analysis: A Cognitive Approach to Cultural Modeling. In: Verma, D. (ed.) Network Science for Military Coalition Operations: Information Extraction and Interaction. IGI Global, Hershey (2010) 3. Wimalasuriya, D.C., Dou, D.: Ontology-based information extraction: An introduction and a survey of current approaches. Journal of Information Science 36(3), 306–323 (2010) 4. Vargas-Vera, M., Motta, E., Domingue, J., Lanzoni, M., Stutt, A., Ciravegna, F.: MnM: ontology driven semi-automatic or automatic support for semantic markup. In: 13th International Conference on Knowledge Engineering and Knowledge Management, Siguenza, Spain (2002) 5. Ciravegna, F., Wilks, Y.: Designing adaptive information extraction for the Semantic Web in Amilcare. In: Handschuh, S., Staab, S. (eds.) Annotation for the Semantic Web. IOS Press, Amsterdam (2003) 6. Miller, G.A., Beckwith, R., Fellbaum, C., Gross, D., Miller, K.J.: Introduction to WordNet: An On-line Lexical Database. International Journal of Lexicography 3(4), 235–244 (2004) 7. Sarawagi, S.: Information extraction. Foundations and Trends in Databases 1(3), 261–377 (2008) 8. Ciravegna, F.: Adaptive information extraction from text by rule induction and generalisation. In: 17th International Joint Conference on Artificial Intelligence, Seattle, Washington, USA (2001) 9. Bontcheva, K., Davis, B., Funk, A., Li, Y., Wang, T.: Human Language Technologies. In: Davies, J., Grobelnik, M., Mladenic, D. (eds.) Semantic Knowledge Management: Integrating Ontology Management, Knowledge Discovery, and Human Language Technologies. Springer, Berlin (2009) 10. Alani, H., Kim, S., Millard, D.E., Weal, M.J., Hall, W., Lewis, P., Shadbolt, N.R.: Automatic Ontology-Based Knowledge Extraction from Web Documents. IEEE Intelligent Systems 18(1), 14–21 (2003)

How Corruption Blunts Counternarcotic Policies in Afghanistan: A Multiagent Investigation Armando Geller, Seyed M. Mussavi Rizi, and Maciej M. L  atek Department of Computational Social Science George Mason University 4400 University Drive Fairfax VA 22030, USA {ageller1,smussavi,mlatek}@gmu.edu

Abstract. We report the results of multiagent modeling experiments on interactions between the drug industry and corruption in Afghanistan. The model formalizes assumptions on the motivations of players in the Afghan drug industry, quantifies the tradeoffs among various choices players face and enables inspection of the time, space and level of supply chain in which one can expect positive and negative impacts of counternarcotic policies. If reducing opium exports is one measure of effectiveness for NATO operations in Afghanistan, grasping the links between corruption and the drug industry should provide a better picture of the second-order interactions between corruption and investment in improving the governance quality, in deploying security forces tasked with eradication and interdiction and in programs to enhance rural livelihoods. Keywords: Afghanistan, Corruption, Multiagent Simulation, Empirical Models, Bounded Rationality.

1

Introduction

In Afghanistan, an illicit economy, a weak government, factionalization and the erosion of norms [1] recur as symptoms of long-lasting conflict along reinforcing divides of modernity versus tradition, urban versus rural, and center versus the periphery. Rampant corruption in a de facto narco-state [2] and a booming drug industry in a de facto highjacked government [3] show two faces of the same Janus guarding the gates to the complex Afghan political economy. In 2009, Afghanistan ranked 179 out of 180 countries in government corruption [4]; 59% of Afghans consider government corruption a greater concern than lack of security and unemployment [5] while bribery is estimated to impose a billion-dollar burden on the Afghan gross domestic product (GDP) annually [6]. Afghanistan has long dominated global opium markets [7] with no less than 30% of the country GDP coming from drug trafficking [5]. Drug money funds both warlords who often double as government officials and the insurgents who fight NATO 

ë

The Office of Naval Research (ONR) has funded this work through grant N00014– 08–1–0378. Opinions expressed herein are solely ours, not of the ONR. We wish to thank Dr. Hector Maletta for providing us with valuable data on Afghan agriculture.

J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, pp. 121–128, 2011. Springer-Verlag Berlin Heidelberg 2011

122

A. Geller, S.M. Mussavi Rizi, and M.M. L  atek

forces and the government. Both sides tax narcotics trade, protect shipments, run heroin labs, and organize farm output in areas they control [8]. After 4 years of refusing to take on the drug industry in Afghanistan [9], the U.S. adopted eradication without alternative livelihoods as the dominant counternarcotic policy in 2006. The policy failed to curb poppy cultivation, but it is credited as “the single biggest reason many Afghans turned against the foreigners” [10]. In 2007, the U.S. sought to complement eradication with rural development programs like the ineffective wheat distribution program in 2008: the program was based on unusually high price ratio of wheat to poppy; failed to address farmers’ incentives for poppy cultivation, and undermined the local seed market. Mindful of the antagonizing effects of eradication on Afghan farmers and convinced of the relative efficiency of interdiction [11,12], in 2009 the Obama administration deemphasized eradication and focused on interdiction and rural development instead [13]. Experts agree that illicit economic activities, especially cultivating poppies and trading narcotics feed government corruption in Afghanistan [14, 15]. Yet, identifying the weaknesses of current counternarcotic efforts in Afghanistan has proven easier than explaining the causal links between corruption and the drug industry; thus outlining effective policies to combat them both. Goodhand [16] finds that counternarcotic policies fuel conflict, because they render the complex interdependencies between rulers and peripheral elites unstable and Felba-Brown [17] blames the same complex interdependencies between economic incentives and local power structures for the eventual failure of the current counternarcotic and anti-corruption policies. We report the results of multiagent experiments on the interaction between the drug industry and corruption in Afghanistan. We have argued in [18] that multiagent modeling is the only approach that can explain historical variations in opium export volumes and prices given the incentives and constraints of Afghan farmers, traders, traffickers and power brokers. Our model formalizes assumptions on the motivations of players in the Afghan drug industry, quantifies the tradeoffs among various choices players face and enables the inspection of the time, space and the level of supply chain in which one can expect positive and negative impacts of counternarcotic policies. On a more applied note, if reducing opium exports is one measure of effectiveness for NATO operations in Afghanistan, grasping the links between corruption and the drug industry should provide a better picture of the second-order interactions between corruption and investment in improved governance, in security forces tasked with eradication and interdiction and in programs to enhance rural livelihoods.

2

The Model: Actors, Behaviors and Validity

Figure 1 outlines the static structure of our model that consists of three basic agent classes of farmers, traders and power brokers. While power brokers can be used to represent ISAF and insurgents, they mainly represent district and province governors who engage in interdiction, eradication or racketeering. Each agent class needs to operate across multiple contexts and has multiple types of decisions to make:

How Corruption Blunts Counternarcotic Policies in Afghanistan

Household

= implements Steppable

Household

District

1..*

123

= extends MasonGeometry

Person 1 boolean sex int age double nonFarmEmployment 80000

1300000 1

Farmer

trades

*

Market

*

monitors 1..* and trades

Trader

1 *

35000 *

3RZHU%URNHU

Village

1 * controls

3000000 1 LandLot double size int landType

100 City int population

350

* 1

*

1 District

Fig. 1. Simplified UML class diagram of the static structure of the model. Main agent types—farmers, traders and power brokers—extend the household class. Therefore, they are located in a specific place and embedded in kinship and social networks. Their welfare and security conditions can also be monitored. Numbers denote approximate counts of each agent class in the model.

• Traders (a) negotiate price, place and size of trades; (b) route shipments, react to interdiction and banditry, corrupt power brokers; and (c) adjust the size of target stock. • Farmers (a) allocate land, decide when to harvest; and (b) rent and offer labor, react to eradication, corrupt power brokers. • Power brokers (a) determine protection rates; (b) respect or renege on mutual agreements; and (c) allocate forces to interdiction and eradication. We faced two challenges in our effort: (a) translating qualitative information into agent purpose, cognition and reasoning and (b) endowing agents with necessary attributes and placing them in the environment. This challenge was amplified by the fact that we had to find a right balance between using historical behaviors coded as if-then-else rules and representing agent adaptation and anticipation driven by first principles, because the purpose of the model is to support policy decisions. For instance, take a farmer’s decision on what crops to cultivate. To make this decision, a farmer needs to account for government and ISAF eradication policies, crop and labor market conditions, family demographics and food security conditions, and availability of seeds and fertilizer. Farmers are endowed with plots of land characterized by land types with specific yields for any combination of potential crops and climate conditions. Some land types are irrigated, so climatic conditions do not influence their yield much. Some crops, for example poppy, are quite resistant to weather conditions regardless of whether they are

124

A. Geller, S.M. Mussavi Rizi, and M.M. L  atek

Price of a kilogram of opium equivalent (city to border) Simulated Real−life data

600

9000

8000

500

7000

Metric tons

USD / kilogram

Total opium harvests 10000

700

400

300

6000

5000

4000

3000

200

2000 100 1000

0 2000

2001

2002

2003

2004

2005

2006

2007

2008

2009

2010

0

2000

2002

2004

2006

2008

Years Price of retail kilogram of wheat (cities and villages)

Total wheat harvests 5000

Thousands of metric tons

1

Simulated Real−life data

0.9

0.8

0.6

4000 3500 3000 2500 2000 1500 1000 500 0

2000

2002

2004

2006

2008

2010

0.5

Total wheat imports 5000

Thousands of metric tons

USD / kilogram

0.7

4500

0.4

0.3

0.2

0.1

0 1999

2000

2001

2002

2003

2004

2005

2006

2007

2008

2009

2010

4500 4000 3500 3000 2500 2000 1500 1000 500 0

2000

2002

2004

2006

2008

2010

Years

Fig. 2. Output validation against historical data for 1999–2009 compiled from [14, 19], corresponding to city- or province-level observations from Nangarhar, Kandahar, Helmand, Badakhshan, Mazari Sharif and Herat for opium, and Kabul, Kandahar, Jalalabad, Herat, Mazari Sharif and Faizabad for wheat on the left panel. Simulation price time series on the left panel are min–max ranges for the same markets. The right panel shows the country-level poppy harvest and wheat imports and harvest for the same period.

cultivated on irrigated land or not. In [20] we proposed and validated an approach to data fusion that helps build all necessary connections between agents and the environment using available data.

How Corruption Blunts Counternarcotic Policies in Afghanistan

2SLXPKDUYHVWDQGHUDGLFDWLRQ 

2SLXPH[SRUWDQGLQWHUGLFWLRQ

















  



1RLQWHUYHQWLRQ SHUVRQQHO SHUVRQQHO



0HWULFWRQVH[SRUWHG

0HWULFWRQVKDUYHVWHG



























,QWHUYHQWLRQ 



125

  



,QWHUYHQWLRQ 













7LPH Fig. 3. Results of interventions with corruption. On the left panel, we use eradication to suppress opium harvests. On the right panel, we target traders with interdiction to suppress opium exports. Intervention is introduced in year 15 and lifted after year 25. Opium losses before reaching border and export are about 10%. Each condition is an average of 10 independent model runs.

The complexity of interdependencies among agents’ decisions is significant. For example, if a farmer with a large family expects imminent drought when poppy prices are high enough, he may decide to forgo grain cultivation and plant only drought-resistant, labor-intensive poppy, hoping to buy wheat from the market with opium income later. Similarly, some farmer households may decide to fallow some of their land and offer accessory labor for farm wages to other households. At the same time, they need to remember that if they choose to cultivate poppy, they will be either exposed to eradication or need to pay protection bribe. [21, 22, 23] and [24] provide detailed accounts of farmers’ crop choice behavior along with their objectives, reasoning mechanisms and information they take into account. Incorporating such lower-level decision factors into our model guarantees that agents can adapt to changing environment plausibly. After instantiating the model at full scale, we ran it in open loop for 10 years to eliminate transients, then introduced historical weather,1 and global wheat prices approximated by the wholesale price of wheat in Pakistan. We also represented the global heroin market and major events in security conditions.2 The resulting dashboard with simulated and historical dynamics compared for the 1999–2009 period is presented on Figure 2. 1 2

We specified monthly, local weather conditions down to the village level, using vegetation indexes in [25] to correlate yields at harvest. We kept the value of all opium and heroin exports from Afghanistan pinned to $2 billion. We also provided a crude estimate of the number of people dealing with eradication and interdiction at the national level.

Metric tons opium produced

126

A. Geller, S.M. Mussavi Rizi, and M.M. L  atek No intervention

8000

25000 personnel

8000

50000 personnel

8000

Intervention

Intervention

7500

7500

7500

7000

7000

7000

6500

6500

6500

6000

6000

6000

5500

5500

5500

5000

5000

4500

4500

5000

Without corruption With corruption

4500 4000

0

5

10

15

20

25

30

35

4000

0

5

10

15

20

25

30

35

4000

0

5

10

15

20

25

30

35

Time

(a) Eradication No intervention

25000 personnel

Metric tons exported

7000

50000 personnel

7000

7000 Intervention

6500

6500

6000

6000

6000

5500

5500

5500

5000

5000

5000

4500

4500

4500

4000

4000

4000

3500

3500

3500

3000

3000

Without corruption With corruption

3000 2500 2000 0

5

10

15

20

25

30

2500 35

2000

Intervention

6500

2500 0

5

10

15

20

25

30

35

2000

0

5

10

15

20

25

30

35

Time

(b) Interdiction Fig. 4. Sensitivity of interventions to corruption. The same experiment design as on Figure 3 has been used. Each condition is an average of 10 independent runs.

3

Policy Experiments

We performed experiments against stationary environmental conditions, with different magnitudes of forces dedicated to either eradication or interdiction after running the model for 15 years. The results of interventions in the presence of corruption are summarized on Figure 3. On Figures 4(a) and 4(b) we repeat the process, this time contrasting the default situation with a universe without corruption. The results show that corruption is merely one mechanism that causes opium production and exports to rebound during and after intervention. Under the eradication regime, prices shift and the spatial distribution of poppy cultivation changes; under the interdiction regime traffickers adapt shipment routes. These adaptive and anticipatory processes provide complementary mechanisms explaining the resilience of the system. In a perfect world, interdiction would be much more cost-effective than eradication, by some 30%. In reality, due to the aforementioned factors, the counternarcotic effects of interdiction diminish faster than those of eradication.

How Corruption Blunts Counternarcotic Policies in Afghanistan

4

127

Conclusions

We have outlined our approach to building an empirically-informed multiagent model of the drug industry in Afghanistan and used it to study the question of resilience of the opium economy to counternarcotic policies implemented by a government riddled with corruption. We have presented the results of model validation against historical conditions and used experiments to quantify the reduction in the effectiveness of eradication and interdiction, due to the presence of corruption of officials tasked with implementing these policies. We have demonstrated that corruption blunts any eradication and interdiction efforts.3 Our model suffers from a number of limitations that open windows for further model development, experiments and field research. On the production side, we have forgone the “salaam” credit system and how eradication compounds farmers’ debt. Our farmers and traders live in a world devoid of existing or emergent social norms: They simply seek survival. On the scenario analysis side, we have made the sweeping assumption that all policy enforcers are subject to corruption. While the generality of this assumption helps us construct “grim scenarios” with conservative estimates, in reality, multiple agencies, some of whom under direct operational supervision of U.S. counternarcotic personnel, carry out operations.

References 1. Rubin, B.: Saving Afghanistan. Foreign Affairs 86(1), 57–78 (2007) 2. Burnett, V.: Crackdown On Afghanistan’s Cash Crop Looms: In War on Drugs. Authorities Seek to Uproot Poppies. Boston Globe, September 18 (2004) 3. Baldauf, S., Bowers, F.: Afghanistan Riddled with Drug Ties; The Involvement of Local as Well as High-Level Government Officials in the Opium Trade Is Frustrating Efforts to Eradicate Poppy Fields. In: Christian Science Monitor, May 13 (2005) 4. Transparency International: Annual Report (2009) 5. United Nations Office of Drugs and Crime: Corruption in Afghanistan. Bribery as Reported by the Victim (2010) 6. Integrity Watch Afghanistan: Afghan Perceptions and Experiences of Corruption. A National Survey (2010) 7. Macdonald, D.: Drugs in Afghanistan: Opium, Outlaws and Scorpion Tales. Pluto Press (2007) 8. Peters, G.: Taliban Drug Trade: Echoes of Colombia. In: Christian Science Monitor, November 21 (2006) 9. Risen, J.: Poppy Fields Are Now a Front Line in Afghanistan War. In: New York Times, May 16 (2007) 10. Hari, J.: Legalize It: Why Destroy Poppies and Afghan Farmers When the World Needs Legal Opiates? In: Los Angeles Times, November 6 (2006) 3

This paper focuses on drugs suppression policies; however, the model includes alternative development, alternative crops, infrastructure development and specific anticorruption policies whose discussions were omitted, because they were not directly relevant to our main findings.

128

A. Geller, S.M. Mussavi Rizi, and M.M. L  atek

11. Rubin, B.R., Sherman, J.: Counter-Narcotics to Stabilize Afghanistan: The False Promise of Crop Eradication. Technical report (2008) 12. Mansfield, D., Pain, A.: Counter-Narcotics in Afghanistan: The Failure of Success? Technical report, Afghanistan Research and Evaluation Unit (2009) 13. Stack, L.: US Changes Course on Afghan Opium, Says Holbrooke. Christian Science Monitor, June 28 (2009) 14. Buddenberg, D., Byrd, W.: Afghanistan’s Drug Industry: Structure, Functioning, Dynamics and Implications for Counter-Narcotics Policy. Technical report, United Nations Office on Drugs and Crime (2009) 15. Katzman, K.: Afghanistan: Politics, Elections, and Government Performance. Technical Report RS21922, Congressional Research Service Report (2010) 16. Goodhand, J.: Corrupting or Consolidating the Peace? The Drugs Economy and Post-conflict Peacebuilding in Afghanistan. International Peacekeeping 15(3), 405– 423 (2008) 17. Felbab-Brown, V.: Testimony before the Senate Caucus on International Narcotics Control, October 21 (2009) 18. Latek, M.M., Rizi, S.M.M., Geller, A.: Persistence in the Political Economy of Conflict: The Case of the Afghan Drug Industry. In: Proceedings of AAAI Complex Adaptive Systems Fall Symposium (2010) 19. Food and Agriculture Organization of the United Nations: Global Information and Early Warning System (2010) 20. Rizi, S.M.M., Latek, M.M., Geller, A.: Merging Remote Sensing Data and Population Surveys in Large, Empirical Multiagent Models: The Case of the Afghan Drug Industry. In: Proceedings of the Third World Congress of Social Simulation (2010), http://css.gmu.edu/projects/irregularwarfare/remotesensing.pdf 21. Maletta, H., Favre, R.: Agriculture and Food Production in Post-war Afghanistan: A Report on the Winter Agricultural Survey 2002-2003. Technical report, Food and Agriculture Organization (2003) 22. Mansfield, D.: What is Driving Opium Poppy Cultivation? Decision Making Amongst Opium Poppy Cultivators in Afghanistan in the 2003/4 Growing Season. In: UNODC/ONDCP Second Technical Conference on Drug Control Research, pp. 19–21 (2004) 23. Kuhn, G.: Comparative Net Income from Afghan Crops. Technical report, Roots of Peace Institute (2009) 24. Chabot, P., Dorosh, P.: Wheat Markets, Food Aid and Food Security in Afghanistan. Food Policy 32(3), 334–353 (2007) 25. National Aeronautics and Space Administration: The Moderate Resolution Imaging Spectroradiometer (MODIS) Data (2010)

Discovering Collaborative Cyber Attack Patterns Using Social Network Analysis Haitao Du and Shanchieh Jay Yang Department of Computer Engineering Rochester Institute of Technology, Rochester, New York 14623

Abstract. This paper investigates collaborative cyber attacks based on social network analysis. An Attack Social Graph (ASG) is defined to represent cyber attacks on the Internet. Features are extracted from ASGs to analyze collaborative patterns. We use principle component analysis to reduce the feature space, and hierarchical clustering to group attack sources that exhibit similar behavior. Experiments with real world data illustrate that our framework can effectively reduce from large dataset to clusters of attack sources exhibiting critical collaborative patterns. Keywords: Network security, Collaborative attacks, Degree centrality, Hierarchical clustering.

1

Introduction

The increasing and diverse vulnerabilities of operating systems and applications, as well as freely distributed cyber attack tools have led to significant volume of malicious activities on the Internet. Hackers can collaborate across the world or control zombie machines to perform Distributed Denial-of-service (DDoS) or other coordinated attacks. One of the common challenges in cyber attack analysis is the effectiveness in analyzing large volume of diverse attacks across the Interent. Ongoing research and practice has focused on investigating attacks via statistical analysis [12] [2], and defining traffic level similarities to cluster Internet hosts [11]. This set of work essentially answers the questions: “what are the most attacked services?” and “what are the most notorious attack sources and their profiles?”. For Botnet attacks, which is a type of collaborative attacks, Gu et al. have proposed to use X-means clustering to analyze TCP/UDP traffic level features [5]. This paper analyzes general collaborative cyber attacks and aims at addressing the questions of “what is the relationship between attack sources?” and “what are the collaborative attack patterns?” by using only the source and target address information in a large dataset. We consider the UCSD Network 

The authors would like to thank The Cooperative Association for Internet Data Analysis (CAIDA) at University of California at San Diego (UCSD) to provide the dataset, the researchers at CAIDA: Claffy, Aben, Alice for early discussions. This work also utilizes Pajek [3], Cytoscape [10] and Network Workbench [9] for analysis.

J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, pp. 129–136, 2011. c Springer-Verlag Berlin Heidelberg 2011 

130

H. Du and S.J. Yang

Telescope dataset [1], which monitors a Class-A network and provides malicious traffic that presumably represents 1/256 of global attacks [8]. We extract the most fundamental information, source and destination address, from a 2-day data collected in Nov. 2008. An Attack Social Graph (ASG) is defined based on the extracted information. Borrowing the idea of degree centrality, six features are derived based on the ASG to describe the relationship among attack sources. Agglomerative hierarchical clustering are utilized to highlight critical attackers and group similar sources. The result shows our methodology can be used to provide abstract information on a group of attack sources and identify possible leader and conspirators from a large dataset.

2

Social Network Modeling of Cyber Attacks

Definition 1 (Attack Social Graph). An Attack Social Graph ASGT (V, E) is a directed graph representing the malicious traffic within a time interval [0, T ], where a vertex v ∈ V is a host, and an edge e(u,v) ∈ E exists if attacks are observed between u and v. Edge direction is from the attacking host to its target. Network Telescope does not respond to any requests, nor do they send any traffic within the Class-A network; Also, it does not collect traffic flowing outside Network Telescope. Therefore, ASGT for the Network Telescope data is a bipartite graph. The vertex set V can be divided into two disjoint sets, Vs and Vt , which represent the set of attack sources and the set of attacked targets, respectively. Moreover, the edge directions can only be from u ∈ Vs to v ∈ Vt . Let din (v) and dout (v) denote the in-degree and out-degree of vertex v, respectively. The degree of vertex in Vs and Vt satisfies: ∀u ∈ Vs , din (u) = 0 and ∀v ∈ Vt , dout (v) = 0. ASGT is comprised of many subgraphs that are disconnected from each other. Each subgraph denotes “local” attacks occurring between Vs ⊂ Vs and Vt ⊂ Vt . The subgraphs vary significantly in size. Figure 1 shows two subgraph examples in ASG5 , where Fig. 1(a) contains 87, 781 vertices1 , and Fig. 1(b) contains 35 vertices. . Definition 2 (Attack Conspirators). Let Tu = {v | e(u,v) ∈ E, ∀v ∈ Vt } be the targets attacked by u ∈ Vs , the attack conspirators of u ∈ Vs , denoted as Cu , . is a set of vertices: Cu = {v | Tu ∩ Tv = ∅, ∀v ∈ Vs }. Consider Fig. 1(b), host B is attacked by 20 sources, which is significantly higher than the average number of attacks, if targets are randomly chosen. Therefore, the sources attacking B might be collaborating or controlled as zombie machines. Similar to B, host C is attacked by a number of sources. Note that host A is a common culprit to attack both B and C which makes A suspicious to controlling all other attacking sources. The hosts who attack the same targets(s) as A does are referred to as the conspirators of A. With a more careful examination, A is the only host in Fig. 1(b), that has a large number of conspirators |CA |, and for most v ∈ CA , dout (v) = 1. This forms the basis of our algorithms that extract suspicious leaders of collaborative attacks. 1

For better visualization, the vertices that has din (v) = 1 are removed in Fig. 1(a).

Discovering Collaborative Cyber Attack Patterns

131

(a) A Large Subgraph of ASG5

(b) A Small Subgraph of ASG5 Fig. 1. Examples of Connected Subgraphs of ASG5

3 3.1

Social Network Analysis of Collaborative Attacks Subgraphs with Collaborative Attacks

ASGT can be too large to be visually analyzed, Table 1 shows the size of ASGT for different T , based on the UCSD Network Telescope dataset. Figure 2(a) plots |Vs |,|Vt | and |E| in log scale. The fact |Vs |  |Vt | indicates a relatively small number of hosts attacking a Class-A network. While ASGT is large, much of it do not contain collaborative attacks. The bipartite subgraphs can be categorized as: one-attacks-one,one-attacks-many, many-attack-one, and many-attack-many cases. Examples of these four types are given in Fig. 3 from left to right. Oneattacks-one and one-attacks-many do not constitute collaborative attacks since Cu = ∅, for all attack sources u in these subgraphs. Table 1 also shows the size of reduced ASGT after removing the one-attacksone and one-attacks-many cases. Figure 2(b) shows the proportion of remaining vertices and edges, i.e., |Vs |/|Vs |, |Vt |/|Vt |, |E  |/|E| for different T values. The reduction seems less effective for a large T : almost 90% of attack sources has at least one conspirator at T = 60. This is due to the higher chance of selecting

132

H. Du and S.J. Yang Table 1. Attack Social Graph Size for Different Time Interval T (minutes)

T 1 3 5 10 30 60

|Vs | 1,685 7,852 15,158 37,342 118,485 203,932

Original ASGT |Vt | |E| 5,818 6,080 43,862 46,498 115,457 122,738 437,780 470,442 2,558,898 3,098,040 5,577,978 7,844,522

(a) Original Graph Size

|Vs | 367 3,457 8,707 27,524 103,144 186,632

Reduced ASGT |Vt | |E  | 109 371 1,118 3,754 2,886 10,167 12,937 45,599 356,098 895,240 1,529,333 3,795,877

(b) Proportion of Remaining Graph Size

Fig. 2. Original Graph Size and Proportion of Remaining Graph Size

Fig. 3. Examples of ASGT Subgraph Patterns

overlapping targets over a longer time interval. To reduce the possibility of analyzing collaborative patterns that contain targets due to random selections, we suggest to analyze ASGT for relatively small T , such as 5 to 10 minutes. This paper considers T = 5 as an example to illustrate our methodology. 3.2

Feature Reduction

The reduced ASGT contains collaborative attacks of different types. Borrowing the idea from social network analysis, collaborative attacks are treated similarly to co-authorship, in which, the authors have a lot of publications with different

Discovering Collaborative Cyber Attack Patterns

133

co-authors are the “leaders” in the research community. We adopt Freeman’s notion of degree centrality measurement [4] to derive six features for each u ∈ Vs to differentiate sources of collaborate attacks in the reduced ASGT . Consider six features are the number the attack sources u ∈ Vs in a reduced ASGT , the  of targets |Tu |, the sum of the targets’ in-degree x∈Tu din (x), the standard deviation of targets’ in-degree σx∈T U din (x), the number of conspirators |Cu |, the sum of conspirators’ out-degree x∈Cu dout (x) and the standard deviation of conspirators’ out-degree σx∈Cu dout (x). These six features are further analyzed using Principle Component Analysis (PCA) [7]. The usage of PCA produces uncorrelated principle components that gives the best separation among the attack sources. As a result, two principle components are found, and the attack source set Vs is further reduced into fewer “feature points” on a 2D plane. Each feature point represents one or more attack sources sharing the same degree centrality features. For example, the ASG5 shown in Table 1 is reduced to 837 feature points, as shown in Fig. 5(a).

4

Clustering Analysis of Collaborative Patterns

The feature points on the 2D plane can be in close proximity and it is logical to perform clustering algorithms to discover similar collaborative patterns. For the reduced ASG5 dataset, most feature points are in cluttered regions on the 2D plane and few are outliers. When zooming into the cluttered region, similar distribution with cluttered regions and few outliers again occurs. This phenomenon can continue for several iterations. In order to better discover the community structure, Agglomerative Hierarchical Clustering [6] is used for analysis. Figure 4 shows the resulting log-scale dendrogram with respect to the reduced ASG5 . The slow and steady staircase climbing exhibited in the left side of the logscale dendrogram suggests that it is challenging to find the community structure from this data. Therefore, as discussed earlier, we investigate the dendrogram and the corresponding feature points by regions. Figure 5(a) shows all the feature points of the reduced ASG5 in the 2D plane and Fig. 5(b) plots the corresponding dendrogram showing only the top-30 branches. Plotting the top-30 branches reveals good community structure, as will be illustrated below. The four branches shown on the left of Fig. 5(b) actually

Fig. 4. The Log Scale (y-axis) Dendrogram of Feature Points on ASG5

134

H. Du and S.J. Yang

(a) Feature Points of Reduced ASG5

(b) Dendrogram of Points for Fig. 5(a)

(c) Feature Points within Cluster A

(d) Dendrogram of Points for Fig. 5(c)

(e) Feature Points within Cluster D

(f) Dendrogram of Points for Fig. 5(e)

Fig. 5. Feature Points and The Corresponding Dendrogram

contains 777 out of the 837 points, i.e., the slow and steady staircase in Fig. 4. These 777 points are clustered as Cluster A and a zoom-in view of Cluster A is shown in Fig. 5(c). Figure 5(d) shows the corresponding dendrogram of feature points within Cluster A. Again only the top branches are chosen to reveal the community structure. The most dense region within Cluster A is Cluster D, and the feature points within Cluster D are further analyzed. Additional iterations can be run to reveal the sub-community structure. Figures 5(e) and 5(f) show the feature points in 10 th iteration and the corresponding dendrogram. By analyzing the clusters and sub-clusters from Fig. 5, we find several sets of interesting collaborative attack patterns. First, the five feature points within Cluster C are actually five distinct attack sources that attack a large number of targets within the Class-A network monitored by the Network Telescope. These five sources are outliers on the 2D plane (see Cluster C in Fig. 5(a)), and a zoomin plot of Fig. 1(a) is shown in Fig. 6(a) to identify these sources in the ASG5 .

Discovering Collaborative Cyber Attack Patterns

(a) Outliers in Reduced ASG5

135

(b) Sample Hosts in Cluster B

(c) Sample Patterns in Cluster H Fig. 6. An Example of Discovered Outliers and Clusters

 Second, all sources in Cluster B have x∈Cu dout (x) > 104 . This is due to that these sources all have a “special” conspirator 125.211.214.160, which is an attack sources in Cluster C and its out-degree is 18, 920. Being a conspirator of such host makes their features significantly different from others and thus form a cluster. Figure 6(b) identifies some of the Cluster B sources in a zoom-in view of ASG5 . Third, Cluster A contains a large number of attack sources that are distinct from Cluster A and B, but do not share much common features. By analyzing the sub-regions within Cluster A, we can reveal some of the collaborative attack patterns. Figure 6(c) shows several examples of “several-attack-one” patterns identified in Cluster H – recall Fig. 5(e) and 5(f).

5

Conclusion and Future Work

Going beyond classical statistical approaches, cyber attack analysis can benefit from social network analysis. This work defined Attack Social Graph to provide an effective representation of relationship among attack sources. Borrowing the idea of degree centrality measurement and agglomerative hierarchical clustering, we discover various types of collaborative attack patterns. Our framework effectively reduces from large dataset and reveals possible leaders of collaborative cyber attacks. Future extension and ongoing work include the use of weighted ASG, where the weights can be defined based on the number of packets exchanged. In addition, a moving time window allows for analysis of how collaborative patterns evolve. Finally, other features, such as geographical and application level information can be used to further differentiate collaborative attacks.

136

H. Du and S.J. Yang

References 1. Aben, E., et al.: The CAIDA UCSD Network Telescope Two Days in November 2008, Dataset (2008), http://www.caida.org/data/passive/telescope-2days-2008_dataset.xml 2. Allman, M., et al.: A brief history of scanning. In: Proceedings of the 7th ACM SIGCOMM Conference on Internet Measurement, p. 82 (2007) 3. Batagelj, V., Mrvar, A.: Pajek-program for large network analysis. Connections 21(2), 47–57 (1998) 4. Freeman, L.: A set of measures of centrality based on betweenness. Sociometry 40(1), 35–41 (1977) 5. Gu, G., et al.: Botminer: clustering analysis of network traffic for protocol andstructure independent botnet detection. In: Proceedings of the 17th Conference on Security Symposium, pp. 139–154 (2008) 6. Jain, A.K., Murty, M.N., Flynn, P.J.: Data clustering: a review. ACM Computing Surveys 31(3), 264–323 (1999) 7. Jolliffe, I.: Principal component analysis. Springer Series in Statistics (2002) 8. Moore, D., Shannon, C., Voelker, G., Savage, S.: Network telescopes: Technical report. CAIDA (2004) 9. NWB Team: Network workbench tool. Indiana University, Northeastern University, and University of Michigan (2006), http://nwb.slis.indiana.edu 10. Shannon, P., et al.: Cytoscape: a software environment for integrated models of biomolecular interaction networks. Genome Research 13(11), 24–98 (2003) 11. Wei, S., et al.: Profiling and clustering internet hosts. In: Proceedings of the International Conference on Data Mining (2006) 12. Yegneswaran, V., et al.: Internet intrusions: Global characteristics and prevalence. In: Proceedings of the International Conference on Measurement and Modeling of Computer Systems, p. 147 (2003)

Agents That Negotiate Proficiently with People Sarit Kraus Dept. of Computer Science Bar-Ilan University, Ramat Gan 52900 Israel and Institute for Advanced Computer Studies University of Maryland, MD 20742 USA [email protected]

Abstract. Negotiation is a process by which interested parties confer with the aim of reaching agreements. The dissemination of technologies such as the Internet has created opportunities for computer agents to negotiate with people, despite being distributed geographically and in time. The inclusion of people presents novel problems for the design of autonomous agent negotiation strategies. People do not adhere to the optimal, monolithic strategies that can be derived analytically, as is the case in settings comprising computer agents alone. Their negotiation behavior is affected by a multitude of social and psychological factors, such as social attributes that influence negotiation deals (e.g., social welfare, inequity aversion) and traits of individual negotiators (e.g., altruism, trustworthiness, helpfulness). Furthermore, culture plays an important role in their decision making and people of varying cultures differ in the way they make offers and fulfill their commitments in negotiation. In this talk I will present the following two agents that negotiate well with people by modeling several social factors: The PURB agent that can adapt successfully to people from different cultures in complete information settings, and the SIGAL agent that learns to negotiate successfully with people in games where people can choose to reveal private information. These agents were evaluated in extensive experiments including people from three countries. I will also demonstrate how agents' opponent modeling of people can be improved by using models from the social sciences.

Acknowledgements. This research is based upon work supported in part by the U.S. Army Research Laboratory and the U.S. Army Research Office under grant number W911NF-08-1-0144 and under NSF grant 0705587.

References 1. Gal, Y., Grosz, B., Kraus, S., Pfeffer, A., Shieber, S.: A framework for Investigating Decision-Making in Open Mixed Networks. Artificial Intelligence Journal 174(18), 1460–1480 (2010) 2. Gal, Y., Kraus, S., Gelfand, M., Khashan, H., Salmon, E.: Negotiating with People across Cultures using an Adaptive Agent. ACM Transaction on Intelligent Systems and Technology (2011) 3. Peled, N., Kraus, S., Gal, K.: A Study of Computational and Human Strategies in Revelation Games. In: Proc. of AAMAS 2011 (2011)

J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, p. 137, 2011. © Springer-Verlag Berlin Heidelberg 2011

Cognitive Modeling for Agent-Based Simulation of Child Maltreatment Xiaolin Hu1 and Richard Puddy2 1

Department of Computer Science, Georgia State University, Atlanta, GA 30303 2 Atlanta, GA 30341

Abstract. This paper extends previous work to develop cognitive modeling for agent-based simulation of child maltreatment (CM). The developed model is inspired from parental efficacy, parenting stress, and the theory of planned behavior. It provides an explanatory, process-oriented model of CM and incorporates causality relationship and feedback loops from different factors in the social ecology in order for simulating the dynamics of CM. We describe the model and present simulation results to demonstrate the features of this model. Keywords: Child maltreatment, cognitive modeling, agent-based simulation, parental efficacy, parenting stress, theory of planned behavior.

1 Introduction Child maltreatment (CM) negatively impacts child development and results in long term health problems as well. Understanding and preventing CM is a challenging task due to the dynamic and complex nature of this phenomenon. The complexity results from a system containing multi-level (individual, relationship, community, societal) factors across the social ecology (Belsky, 1980), diversity of actors (such as families, schools, government agencies, health care providers) that potentially affect maltreatment, and multiplicity of mechanisms and pathways that are not well studied or understood. Complex systems science and agent-based modeling offer tremendous promise in this area because they have proven to be a powerful framework for exploring systems with similar characteristics (Hammond, 2009). Using a complexity science-informed approach, in previous work (Hu and Puddy, 2010) we developed an agent-based model (ABM) for studying CM and CM prevention. The ABM uses a resource-based conceptual model to simulate the occurrence of CM. It models a community of agents, each of which corresponds to a family unit and includes a parent-child relationship. CM occurs when there exists unmet child need, i.e., the parent care is insufficient to meet the child need. Agents are connected to each other through social network, which acts as a potential resource for agents to obtain support in providing child care. The developed ABM focuses on the community level of the social ecology. It was able to simulate the impacts of several community level factors such as the density of social connection, community openness, and community resource on the rate of CM (Hu and Puddy, 2010). J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, pp. 138–146, 2011. © Springer-Verlag Berlin Heidelberg 2011

Cognitive Modeling for Agent-Based Simulation of Child Maltreatment

139

While the previous work built a foundation for applying agent-based modeling and simulation to studying CM, extensions are needed in order to adequately model the dynamics of CM. In particular, in the previous work each agent itself is pretty simple in the sense that it only models the resource aspect to account for the occurrence of CM. Although resource is an important factor in CM, the occurrence of CM is a result of a cognitive process that impacts parents’ decision to engage in aggression toward children (Milner, 1993, 2000). Thus in order to adequately model the dynamics and the complexity of CM it is necessary to build a cognitive process into the agent model. Motivated by this need, this paper presents cognitive modeling of CM. The developed cognitive model is inspired from several sources of cognitive science and CM psychology, including the Theory of Planned Behavior (TPB) (Ajzen, 1985; Ajzen, 1991), self-efficacy theory (Bandura, 1977; Bandura, 1994), and models of parenting stress (Belsky, 1984; Hillson and Kuiper, 1994). TPB postulates three conceptually independent determinants of behavioral intention: attitude toward the behavior, subjective norm, and the perceived behavioral control. Among them the perceived behavior control is originated from the concept of self efficacy. For the phenomenon of CM, a specialized type of self-efficacy, parental efficacy, can be defined as the extent to which parents believe they can influence the context in which their child is growing (Shumow and Lomax 2002). This concept of parental efficacy is used in our model to influence the perceived behavior control in TPB. A brief introduction of these theories/models is omitted here and can be found in an extended version of this paper (Hu and Puddy, 2011). Our work of cognitive modeling aims to provide an explanatory, process-oriented model of CM, incorporating causality relationship and feedback loops from different factors in the social ecology of CM. It is hoped that the developed model will support more accurate agent-based simulation of CM and provide more accountable insights for preventing CM. It is important to note that there are different types of child maltreatment, including physical abuse, emotional abuse, sexual child abuse, and neglect. While these different types of maltreatment may follow different cognitive pathways, in this work we are less concerned about the specific type of maltreatment. Instead, as a simulation model aiming towards a computational analysis of the phenomenon of CM, this work pays more attention to a general structure of the cognitive process of CM, regardless its specific format. Because of this, in our paper we use abstract terms “responsive to child need” and “not responsive to child need” to describe parents’ behavior, where “not responsive to child need” means maltreatment. With that said, however, we note that the development of our model is mainly informed from the physical abuse and neglect types of maltreatment, and thus may not be meaningful for the other maltreatment types. Also note that this paper focuses on the modeling of a single agent. Multi-agents interactions and community level factors can be added later (see, e.g., Hu and Puddy, 2010) and are not explicitly considered in this paper.

2 Cognitive Modeling of Child Maltreatment In developing agent’s cognitive model of child maltreatment, we employ TPB as a basic mechanism to compute the behavioral intention of the agent (regarding being responsive or not responsive to child need). A significant portion of the cognitive model deals with how to compute the perceived behavior control used in TPB. In our

140

X. Hu and R. Puddy

work, the perceived behavior control is the result of parental efficacy modulated by the stress level. Fig. 1 shows the major elements and their relationships of the cognitive model, followed by the description of how the model works. Due to space limitation, the description focuses on the computational aspect by showing how the major elements are computed and updated. Corresponding mathematical equations are omitted in this paper, and can be found in the Appendix of an extended version of this paper (Hu and Puddy, 2011). We note that the purpose of this paper is to explore the possible dynamics of CM without an attempt to have fidelity to real world observations (real world fidelity and validation will be attempted in later phases). Because of this, some parameter values (e.g., coefficients, weights) used in the equations are empirically determined based on what we think are reasonable within the context of CM. Due to the same reason, we simplify the quantification of all the elements of our model (except for parenting stress and contextual stress) by representing them on an arbitrary scale of 0-100. The units of the scale are purposely not specified in order to simply represent relative values of the variables under study. Also note that in the following description, elements such as child need, family resource, and community (with social network) have been defined in previous work (Hu and Puddy, 2010) and thus are used in this paper with minimal explanations. social normative beliefs /social norm

child need

community

resource-based efficacy resources

family resource

perceived community resource

attitude

w2 perceived behavioral control

parental efficacy experiencebased efficacy

stress level

parenting stress

w3

no

w1 intention

responsive to child need?

contextual stress

yes no

community unmet child need >0 feedback

enough resource? yes

unmet child need =0

Fig. 1. Cognitive modeling of child maltreatment (of a single agent)

Parental Efficacy An agent’s parental efficacy defines the agent’s belief about its capability of taking care of the child (i.e., meeting the child’s need). In our work, an agent’s parental efficacy is formed from two sources: the resource-based efficacy and the experiencebased efficacy. The resource-based efficacy captures the portion of efficacy induced from the agent’s resource compared to the child need. It thus depends on the

Cognitive Modeling for Agent-Based Simulation of Child Maltreatment

141

difference between the resource and child need – the larger the difference, the higher the resource-based efficacy. In our model, this relationship is captured by a linear equation between the resource-based efficacy and the difference between resource and child need. Note that an agent’s resource can include both family resource and perceived community resource. The later depends on community level factors such as social network and community resource (Hu and Puddy, 2010). The experience-based efficacy captures the portion of efficacy induced from the agent’s past experience of taking care of the child. The experience-based efficacy changes over time and is dynamically updated based on the agent’s experience of meeting the child need. In general, if the agent successfully meets the child need, its experience-based efficacy increases; otherwise, its experience-based efficacy decreases. In our model, the experience-based efficacy increases or decreases in a predefined rate based on whether the child need is met or not. A coefficient is used to model the ceiling effect when increasing experience-based efficacy: as the experience-based efficacy becomes larger, the increase rate of the experience-based efficacy becomes slower. The parental efficacy is the weighted sum of the resource-based efficacy and experience-based efficacy. We assign a larger weight for the experience-based efficacy because we consider parent-child interactions as routine events. Under routine circumstances, an individual may well utilize previous performance level (obtained from experience) as the primary determinant of self-efficacy (Gist and Mitchell, 1992). Stress Level An agent’s stress level comes from two different sources: contextual stress and parenting stress. Belsky (1984) has suggested that there are three contextual sources of stress: the marital relationship, social networks, and employment that can affect parenting behavior. In our work, these stresses are modeled by a single entity called contextual stress and is assumed to be unchanged during the simulation. Besides the contextual sources of stress, parenting stress (i.e., stress specific to the parenting role) also directly and indirectly affected parenting behavior (Rodgers, 1998). Parenting stress results from the disparity between the apparent demands of parenting and the existing resources to fulfill those demands (Abidin, 1995). In our model, the parenting stress is due to being a parent for satisfying the child need. It is dynamic in nature as described below: 1) when the amount of parent care provided by the agent is much less than the family resource (this means, e.g., the agent can “easily” satisfy the child need), the parenting stress decays over time– it decreases from the previous value; 2) otherwise when the amount of parent care is close to the full capacity of family resource (this means, e.g., the agent needs to exert its full effort in order to satisfy the child need), the parenting stress accumulates over time – it increases from the previous value. In our model, we use a threshold ratio 80%: when more than 80% of the family resource is used, parenting stress increases; otherwise the parenting stress decreases. Also note that in the current model, when an agent chooses “not responsive to child need”, it uses zero of its family resource (as in neglect). Thus in this case, the parenting stress decreases because the agent is not engaged in the parenting behavior. An agent’s stress level is the sum of its contextual stress and parenting stress.

142

X. Hu and R. Puddy

Behavioral Intention Based on TPB, an agent’s behavioral intention is calculated from attitude, social norm, and perceived behavior control (denoted as PBC), multiplying by their corresponding weights. The attitude indicates the agent’s general tendency or orientation on the parenting behavior. The social norm reflects the social context of the agent, and can be computed from the agent’s social network (e.g., the average of all neighbors’ attitudes). The PBC is the result of parental efficacy modulated by the stress level. In our model, the modulation is modeled by a stress-based coefficient that modulates the parental efficacy into the PBC. In general, the higher the stress level, the smaller the stress-based coefficient. In computing the behavioral intention, three weights are used to define the relative percentages of contribution of the three elements: attitude, social norm, and PBC in computing the intention. In our implementation, these weights are determined based on the level of PBC: when an agent has high level of PBC, the attitude, social norm, and PBC have about the same weight; as the level of PBC drops, the weight for the PBC becomes larger. This approach, to some extent, incorporates the idea of “hot” cognition (Abelson, 1963), which is a motivated reasoning phenomenon in which a person's responses (often emotional) to stimuli are heightened. Based on its attitude, social norm, and PBC, an agent’s behavioral intention is the weighted sum of them. This intention determines the agent’s probability of choosing the “responsive to child need” behavior. Thus the higher the intention level, the more likely the agent will be engaged in the parenting behavior for meeting the child need. Finally, if the “responsive to child need” behavior is chosen, the agent will check if it actually has the resource (family resource plus community resource) to meet the child need. Only when this is true, the child need is successfully satisfied. Otherwise (either if the “not responsive to child need” is chosen or there is insufficient resource), there exists unmet child need, i.e., child maltreatment. Simulation of the agent model runs in a step-wise fashion. In every step, the parental efficacy is computed from the current experience-based efficacy and resourcebased efficacy (based on the difference between resource and child need). The stress level is also computed from the contextual stress and the current parenting stress. Then the perceived behavior control is calculated, and the behavioral intention is calculated to determine the probability of choosing “responsive to child need” behavior. Finally, due to the feedback loops, the experience-based efficacy is updated based on if the child need is met or not, and the parenting stress is updated based on the parent care to family resource ratio.

3 Simulation Results We carried out a series of simulation experiments to illustrate the fundamental features of our cognitive model and see what insights we might gain by them. Our experiments aim to provide qualitative understanding of the model dynamics, i.e., to explore the potential directional impact on CM rates when efficacy, stress, and family resource are varied. To do this, we fix the other model parameters as follows: attitude =90, social norm =90, child need =50, perceived community resource =0, initial value of parenting stress =0. The value we chose for each of these elements was designed to reflect a rough cut at the relative contribution of that element in relation to the others.

Cognitive Modeling for Agent-Based Simulation of Child Maltreatment

143

Results in this section are based on simulations using a single agent. Due to space limitation, the result of family resource is not shown in this paper. Our first experiment shows how the initial value of experienced-based efficacy impacts the CM result. In this experiment, we set a high contextual stress 20 (66.7 in a 100 scale), and set the family resource to be 55, which is enough to meet the child need (=50) but still results in parenting stress because it is not significantly larger than the child need. Fig. 2(a) shows how the parenting stress, total stress level (parenting stress plus contextual stress), experience-based efficacy (denoted as EBE), perceived behavior control (denoted as PBC), and the intention level change over time for 2000 time steps when the EBE starts with an initial value of 30. Fig. 2(b) shows the same information when the EBE starts with an initial value of 40. The two figures show two clearly different behavior patterns due to the different initial value of EBE. In Fig. 2(a), since the initial EBE of 30 is relatively small, the agent’s intention of choosing “responsive to child need” is not high enough, leading to significant number of CM. The large number of CM in turn decreases the EBE further, resulting in an adverse self-reinforcing loop that makes EBE rapidly decreases to zero. On the contrary, in Fig. 2(b) the initial EBE of 40 was able to make the agent have high enough intention so the experience of meeting the child need becomes dominant. This in turn increases the agent’s EBE further, resulting in a favorable self-reinforcing loop that gradually raises EBE, PBC, and the intention to a high level. Note that as the agent becomes more and more engaged in the parenting behavior, the parenting stress gradually increases too. The influence of parenting stress will be studied in the next experiment. Initial Experience Based Efficacy (EBE) = 30 parentingStress

totalStress

EBE

PBC

Initial Experience Based Efficacy (EBE) = 40

Intention

parentingStress

100 90

totalStress

EBE

PBC

Intention

100 90 80 70 60 50 40 30 20 10 0

80 70 60 50 40 30 20 10 0 0

200

400

600

800

1000

1200

1400

1600

1800

Time Step

(a) Initial experience-based efficacy=30

2000

0

200

400

600

800

1000

1200

1400

1600

1800

2000

Time Step

(b) initial experience-based efficacy=40

Fig. 2. Dynamic behavior for different initial experience-based efficacy values

To further see the impact of the initial EBE, we vary the initial EBE from 10 to 100 (increase by 10 each time) and carry out the simulations. Fig. 3 shows the dynamic change of EBE over time for different initial values. As can be seen, when the initial value is 30 or smaller, the EBEs all decrease to 0 due to the adverse self-reinforcing loop explained above. When the initial value is 40 or more, the EBEs either gradually increase or gradually decrease and eventually all converge to the same value (about 72). This experiment reveals two important properties of the system impacted by EBE. First, the initial EBE can play a “critical” role in defining the overall experience of parenting – a relative small change in the initial EBE (e.g., from 30 to 40) may result in significantly different system behavior. Second, regardless the initial values,

144

X. Hu and R. Puddy

Dynamic Change of Experience-based Efficacy Over Time (Family Resource = 55; Contextual Stress = 20) 100 90 80

Initial EBE=10

70

Initial EBE=20

60

Initial EBE=30

50

Initial EBE=40 Initial EBE=50

40

Initial EBE=60

30

Initial EBE=70

20

Initial EBE=80

10

Initial EBE=90 Initial EBE=100

0 0

200

400

600

800

1000

1200

1400

1600

1800

2000

Time Step

Fig. 3. The impact of different initial experience-based efficacy values Average model outputs under different contextual stresses (Family Resource = 55; Initial Efficacy = 35) 100 90 80 70 60

parentingStress

50

totalStress

40

Intention

30

Parent_care

20

#CM (x10)

10 0 0

1

2

3

4

5

6

7

8

9

10 11 12 13 14 15 16 17 18 19 20

Contextual Stress

Fig. 4. The impact of contextual stress

eventually the system always converge to one of the two relatively stable states, that is, to two attractors. This type of bifurcation (convergence of the system on one of two relatively stable states) is common in complex systems. Observing this phenomenon here points to the need for further exploration of this system in future work. Our second experiment tests the impact of the contextual stress on the CM result. In this experiment, we set the family resource to be 55, the initial experience-based efficacy to be 35. We vary the contextual stress from 0 to 20 (increase by 1 each time) and run the simulations for 2000 time steps. For each simulation we ran the model 10 times and averaged the results which we present in Fig. 4. Since it takes time for the system to reach a stable state as shown above, we only measure the results between time 1000 and time 2000, and then compute the average of the 1000 time steps. Fig. 4 shows the averaged parenting stress, total stress, intention, parent care, and number of CM. Note the parent care is the total amount of parent care averaged over the 1000 time steps, and the number of CM is the total number of times when there are unmet child need averaged over the 1000 time steps. From the figure we can see that the contextual stress has important impacts on CM. In most situations, when the contextual stress reduces, the number of CM decreases too. However, counter intuitively, this is not always true, as in Fig. 4 when the contextual stress reduces from 18 to 14, the number of CM maintains at the same level. This is because the reduction of

Cognitive Modeling for Agent-Based Simulation of Child Maltreatment

145

contextual stress causes the increase of parenting stress (as shown in Fig. 4) due to the increase of intention. Together they make the overall stress level stay at a constant level. It should be noted that this result does not mean there is no benefit to reduce the contextual stress from 18 to 14 – as the contextual stress is reduced further to below 14, the number of CM decreases too. This experiment shows that the interplay between contextual stress and parenting stress can cause complex, and sometimes counter-intuitive, system behavior.

4 Conclusions This paper extends previous work to develop cognitive modeling for agent-based simulation of CM. The developed model is inspired from the theory of planned behavior, self-efficacy theory, and parenting stress models. Preliminary results using a single agent are presented. Future work includes more analysis and tests of the cognitive model, adding community level factors to carry out multi-agent simulations, and aligning the model parameters with real world data and model validation. Acknowledgments. This work was supported in part by a CDC-GSU seed grant. We thank Charlyn H. Browne and Patricia Y. Hashima for their valuable inputs in developing the model.

References 1. Abelson, R.P.: Computer simulation of ”hot cognition”. In: Tomkins, S.S., Messick, S. (eds.) Computer Simulation of Personality, pp. 277–302. Wiley, New York (1963) 2. Abidin, R.R.: Parenting Stress Index: Professional Manual, 3rd edn. Psychological Assessment Resources, Inc., Odessa (1995) 3. Ajzen, I.: From intentions to actions: A theory of planned behavior. In: Kuhl, J., Beckman, J. (eds.) Action-control: From Cognition to Behavior, pp. 11–39. Springer, Heidelberg (1985) 4. Ajzen, I.: The theory of planned behavior. Organizational Behavior and Human Decision Processes 50, 179–211 (1991) 5. Bandura, A.: Self-efficacy: Toward a unifying theory of behavioral change. Psychological Review 84, 191–215 (1977) 6. Bandura, A.: Self-efficacy. In: Ramachaudran, V.S. (ed.) Encyclopedia of Human Behavior (Vol, pp. 71–81. Academic Press, New York (1994); Friedman, H. (ed.): Encyclopedia of mental health. Academic Press, San Diego (1998), http://www.des.emory.edu/mfp/BanEncy.html (last accessed: 11/05/2010) 7. Belsky, J.: Child maltreatment: An ecological model. American Psychologist 35(4), 320– 335 (1980) 8. Belsky, J.: The determinants of parenting: A process model. Child Development 55, 83–96 (1984) 9. Hillson, J.M.C., Kuiper, N.A.: A stress and coping model of child maltreatment. Clinical Psychology Review 14, 261–285 (1994) 10. Hu, X., Puddy, P.: An Agent-Based Model for Studying Child Maltreatment and Child Maltreatment Prevention. In: Chai, S.-K., Salerno, J.J., Mabry, P.L. (eds.) SBP 2010. LNCS, vol. 6007, pp. 189–198. Springer, Heidelberg (2010)

146

X. Hu and R. Puddy

11. Hu, X., Puddy, P.: Cognitive Modeling for Agent-based Simulation of Child Maltreatment, extended paper (2011), http://www.cs.gsu.edu/xhu/papers/SBP11_extended.pdf 12. Gist, M.E., Mitchell, T.R.: Self-Efficacy: A Theoretical Analysis of Its Determinants and Malleability. The Academy of Management Review 17(2), 183–211 (1992) 13. Milner, J.S.: Social information processing and physical child abuse. Clinical Psychology R&w 13, 275–294 (1993) 14. Milner, J.S.: Social information processing and child physical abuse: Theory and research. In: Hansen, D.J. (ed.) Nebraska Symposium on Motivation, vol. 45, pp. 39–84. University of Nebraska Press, Lincoln (2000) 15. Shumow, L., Lomax, R.: Parental efficacy: Predictor of parenting behavior and adolescent outcomes. Parenting, Science and Practice 2, 127–150 (2002)

Crowdsourcing Quality Control of Online Information: A Quality-Based Cascade Model Wai-Tat Fu and Vera Liao Department of Computer Science, University of Illinois at Urbana-Champaign Urbana, IL 61801, USA wfu/[email protected]

Abstract. We extend previous cascade models of social influence to investigate how the exchange of quality information among users may moderate cascade behavior, and the extent to which it may influence the effectiveness of collective user recommendations on quality control of information. We found that while cascades do sometimes occur, their effects depend critically on the accuracies of individual quality assessments of information contents. Contrary to predictions of cascade models of information flow, quality-based cascades tend to reinforce the propagation of individual quality assessments rather than being the primary sources that drive the assessments. We found even small increase in individual accuracies will significantly improve the overall effectiveness of crowdsourcing quality control. Implication to domains such as online health information Web sites or product reviews are discussed. Keywords: Web quality, user reviews, information cascades, social dynamics.

1 Introduction The rapid growth of the WWW has led to increasing concern for the lack of quality control of online information. Many have suggested that quality assessments of online information could be “crowdsourced” to the public by harnessing the collective intelligence of online users by allowing them recommend contents and share the recommendations with others. Although this kind of crowdsourcing seems to work well in many domains [2, 3, 9], recent studies, however, show that aggregate behavior is often subject to effects of cascades, in which the social dynamics involved during the accumulation of recommendations may not be effective in filtering out low-quality information [8]. Indeed, previous models show that when people imitate choices of others, “bad cascades” may sometimes make low-quality information more popular than high-quality ones [1]. Previous cascade models often assume that users imitate choices of others without direct communication of quality information. We extend previous models by assuming that users can make multiple choices among web sites, and they can communicate their quality assessments to other users. Our goal is to understand how individual quality assessments may moderate effects of information cascades. We investigate how cascades may be influenced by factors such as accuracies of individual quality assessments, confidence levels of accepting other users’ recommendations, and impact of J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, pp. 147–154, 2011. © Springer-Verlag Berlin Heidelberg 2011

148

W.-T. Fu and V. Liao

aggregate user recommendations on choices. The goal is to simulate effects at the individual level to understand how the social dynamics may lead to different aggregate patterns of behavior. Specifically, we investigate how individual quality assessments propagate in a participatory Web environment, and whether information cascades may moderate the effectiveness of aggregate user recommendations, such as whether they may lead to “lock in” to low-quality information; and if so, to what extent will they impact the overall effectiveness of crowdsourcing quality control of information.

2 Background 2.1 Quality Assessment of Online Information While research has shown that people are in general good at choosing high quality Web sites, it is also found that people often browse passively by relying on what is readily available. For example, health information seekers are found to rely on search engines to choose online sources, and may utilize surface features (such as layout and structures of web pages) to infer quality [5]. Accuracies of quality assessments are also found to vary among individuals, and can be influenced by various factors such as background knowledge, Internet experience, cognitive resources available, etc [4]. Liao & Fu [7] also found that older adults, who tended to have declined cognitive abilities, were worse in assessment quality of online health information. Research in the field of e-commerce shows that user reviews can often significantly influence choices of online information, and their impact may be biased by “group opinions” while ignoring their own preferences [3]. Research also shows that effects of cascades and social dynamics may influence effectiveness of user reviews [8] in guiding users to choose high-quality products. However, to our knowledge, none has studied how cascades may be influenced by accuracies of quality assessment at the individual level, which may lead to different emergent aggregate behavioral patterns. In fact, research has shown that in certain situations, even random accumulation of user recommendations may induce more users to follow the trend, creating a “rich gets richer” cascading effect that distorts the perception of quality. On the other hand, previous cascade models often do not allow users to pass their quality assessments to other users, and it is possible that the passing of quality information may moderate effects of cascades. The goal of the current model is to test the interactions between individual and aggregate quality assessments to understand the effectiveness of collective user recommendations on promoting high quality information. 2.2 Cascade Models Many models of information cascades have been developed to understand collective behavior in difference contexts, such as information flow in a social network [6], choices of cultural products in a community [1], or effectiveness of viral marketing in a social network [10]. In the model by Bikhchandani et al [1], information cascades are modeled as imitative decision processes, in which users sequentially choose to accept or reject an option based on their own individual assessment and the actions of other users. Each user can then observe the choice made by the previous users and infer the individual assessments made by them. In most information cascade models,

Crowdsourcing Quality Control of Online Information

149

the users can only observe the choices made by previous users and learn nothing from the users about the quality of the options. Given this limited information, a string of bad choices can accumulate, such that a cascade might lock future users on a lowquality option. Similarly, cascade models of information flow aim at characterizing how likely information will spread across a social network [6, 10] without any specific control on how likely one may assess the quality of the information and how the passing of the quality information to other users may influence how it spreads to other people in the network. It is therefore still unclear how the passing of quality assessments by users, in addition to their choices, might influence information cascades, and how the extent to which the aggregate quality assessments can be relied on to filter out low-quality information.

3 The Model The general structure of the model is to have a sequence of users each deciding whether to select one of the 20 sites (or select nothing). When user i selects a site j, a recommendation will be given to the site based on the user’s imperfect quality assessment. User i+1 will then interpret the recommendation by user i with a certain level of confidence to determine whether to select the same site j or not. If user i+1 does not select j, she will either select a site from a recommended list of sites based on aggregate recommendation by previous users or select nothing. The model therefore aims at capturing the interaction of “word of mouth” information from the neighboring users and the global assessment of quality by previous users. 3.1 The Web Environment We simulated the environment by creating sets of 20 Web sites, with each set having different proportions of sites with good and bad quality. A user can freely pick any site (or not pick any) and give a positive or negative recommendation (or no recommendation) to it. This recommendation is available for the next user, acting as a local signal (i.e., word-of-mouth information) collected from neighboring users. Recommendations will also be accumulated and used to update the list of most recommended sites, which will be available for the next users. The choice of sites and recommendation are therefore influenced by three factors: (1) the perceived quality of the site by the user (the private signal), (2) the recommendation provided by the previous user (the local signal), and (3) the accumulated list of most recommended sites (the global signal). 3.2 A Quality-Based Cascade Model We simulated the sequential choice and recommendation process by a set of 1000 users as they chose and gave recommendation to the set of 20 Web sites. When the simulation began, a user was randomly assigned to a site. The user then evaluated the site, and based on the user’s perceived quality, it will assign a positive or negative recommendation to the site. Specifically, it was assumed that when the site had good quality, there was a probability Q (or 1-Q) that the user would perceive that it had good (or bad) quality. When the perceived quality was good (or bad), there was a

150

W.-T. Fu and V. Liao

probability R (or 1-R) that the user would give a positive (or negative) recommendation. The next user could then interpret the recommendation given by the previous user to infer the quality of the site. If the recommendation was positive (or negative), there was a probability C (or 1-C) that the next user would choose (or not choose) the same site. If the user did not choose the same site, there is a probability L (or 1-L) that it would randomly select from the top recommended list (or not choose any site). Fig. 1 shows the decision tree of the model.

Fig. 1. The quality-based cascade model

3.3 Choice Accuracies Fig. 2 shows the mean proportions of choice of high quality sites by the users with different values of Q (0.5, 0.55, 0.6) in a High (20% low), Medium (50% low), and Low (80% low) quality environments, with (a) low (C=0.5) and (b) high (C=0.9) confidence levels. In the “Top1” environment, when users decided to select from the recommendation list, she always picked the site with the highest recommendation; in the “Top10” environment, the user randomly picked a site among the best 10 sites. Results show significant interactions among the private, local, and global signals. Good individual quality assessments (i.e., Q) in general lead to better overall accuracies in choosing high quality sites, but its effect on choice is strongly magnified by both the global signal (i.e., the recommendation list) and the confidence level in neighboring users’ recommendation (i.e., C). Even when the confidence level is low (C=0.5), good quality assessments can be propagated through the aggregate recommendation list, assuming that other users will select the top recommended sites (in the Top1 environment). However, when users randomly selected sites from a longer list (Top10 environment), this channel of propagation diminished quickly, as shown by the low accuracies in the Top10 environment (accuracies in Top5 were somewhere between Top1 and Top10). When the confidence level was high (C=0.9), good quality assessments can be propagated through “word of mouth” local information, which improved the overall accuracies even in the top-10 environment. In general, even a small increase of Q from 0.5 (random choice) to 0.55 will lead to significant increase in overall choice of the correct Web sites for the aggregate. Results demonstrate how local and global signals can magnify individual assessment of quality to increase the overall effectiveness of crowdsourcing quality control. 3.4 Effects of Information Cascades Fig. 3 shows examples of the simulations that illustrate whether information cascades occur in different conditions. Each figure shows the aggregate (sum of positive and negative) recommendations of the 20 web sites (different color lines) in the Medium

Crowdsourcing Quality Control of Online Information

151

Fig. 2. Choice accuracies of high quality sites by users with different quality assessment accuracies (Q=0.5, 0.55, 0.6) and with low (a) and high (b) confidence levels (C) of other’s recommendation, in environments with High (20%), Medium (50%) and Low (80%) ratio of low quality sites. Increasing Q from 0.5 to 0.55 leads to sharp increase in performance, especially in the Top1 environment and when C is high.

quality environment. Information cascades occur in all conditions, except when Q=0.5, C=0.5 in the Top10 environment, in which apparently there is no clear direction given by each user. In contrast, in the Top1 environment (similar results obtained for different values of C), information cascades do occur through the selection of the most recommended sites. When Q=0.5, because quality assessments are at chance level, information cascades occur whenever successive users choose and assign a positive recommendation on a random site and made the site the most recommended, which was enough to drive other users to keep selecting it. Higher Q value in Top1 environment shows even stronger information cascade, as consistent positive recommendations quickly lock on to a high quality site. In the Top10 environment, when C=0.9, information cascade occur through the propagation of local “word-of-mouth” information. However, when Q=0.5, information cascade occurs much slower, as it requires successive random assessment of high quality of a particular site, but even when that happens, the effect is weak and tends to fade away quickly, as there is no reinforcement by the global signal. However, when Q is high, information cascades make the high quality sites quickly accumulate positive recommendation and increase their chance of being selected from the top-10 list. In other words, the pattern of results suggests that, when quality assessment is accurate, high quality sites tend to stand out quickly from the low-quality ones. Growth in positive recommendation for high quality sites quickly reinforces the flow of quality assessments by individuals. However, cascades themselves do not drive quality assessments to a sustainable level unless individual assessments of quality are low, and even so it requires a relatively rare sequence of random events to induce the cascades. 3.5 Effects of Initial Conditions The above simulations assumed that all sites received no recommendation initially and users randomly choose among them to recommend to others. This may be different from actual situations as “word-of-mouth” information is often distributed

152

W.-T. Fu and V. Liao

Fig. 3. Recommendations (y-axis) given to the sets of Web sites sequentially by 1000 users (x-axis) in the Medium environment. Color lines represent the net aggregate recommendations of 20 Web sites. Cascades occur when one (or more) Web site dominates others as more user recommendations are added. Except when Q=0.5, cascades tend to favor high-quality sites. unevenly, or is subject to viral marketing as people are promoting their Web sites. We therefore simulated the effects of having positive recommendation on a subset of low quality sites and see how well these recommendations could hold up across time. Fig. 4 shows the simulation results in the Medium quality environment with 2 of the low quality sites started off with 20 positive reviews (a higher number simply takes more cycles (users) to reach the same patterns). Consistent with previous simulations, when Q=0.5, the initially positively recommended low quality sites never come down, as users are not able to pass useful quality assessments to other users. In fact, bad cascades tend to drive up these low-quality sites. However, when Q is slightly higher (slightly more than random chance), recommendations of these sites all come down quickly as users collectively give negative recommendations on them. In the Top1 environment, sequential assignments of negative recommendations by multiple users quickly bring down the recommendations for the low quality sites and remove them from the top-1 list. When this happens, positive recommendations accumulate quickly on a high-quality site. In the Top10 environment, even when Q=0.5, when C=0.9, word-of-mouth information may accumulate and lead to cascades, although it is equally likely that the “winning” site has high or low quality. When Q is high, however, recommendations on low-quality sites come down quickly, while high-quality sites accumulate positive recommendations and become more likely to be selected once they appear on the top-10 list. This effect again was magnified by a higher C value, showing that both local and global information significantly reinforces the collective quality assessments by multiple users.

Crowdsourcing Quality Control of Online Information

153

Fig. 4. Recommendations given to the sets of Web sites sequentially by 1000 users in the Medium quality environment. Two of the low-quality sites had 20 positive reviews when the simulation started. When Q=0.5, these low-quality sites tend to stay high, and “bad” cascades often drive them even higher; When Q=0.6, these low-quality sites quickly come down, and “good” cascades tend to magnify aggregate recommendations of high-quality sites.

4 Conclusions and Discussion We extend previous cascade model to understand how quality assessments of information may be magnified by the effects of cascades at both the local (“word-ofmouth”) and global (aggregate recommendation list) levels. We found that in general, even when individual assessments of quality is only slightly better than chance (e.g., p=0.55 or 0.6 that one can correct judge quality), local and global signals can magnify the aggregated quality assessments and lead to good overall quality control of information, such that users can more likely find high quality information. We also show that when users can pass quality assessments to others, cascades tend to reinforce these assessments, but seldom drive the assessments to either direction. The results at least partially support the effectiveness of crowdsourcing quality control: even when users are far from perfect quality assessment, so long as the overall quality assessments can flow freely in the social environment, effects of the aggregated quality assessments can be magnified and are useful for improving the choice of high-quality Web sites. The current model allows users to choose among many options (or choose nothing), which seems to make cascades less likely to occur (compared to a binary choice), but also more unpredictable. The mixing of information cascades of many scales among competing choices results in a turbulent flow to outcomes that cannot be easily predicted [8], but requires a model to predict the multiplicative effects among different signals. Our model shows that cascades are not necessarily bad, they could be effective forces that magnify the aggregated effort of quality assessments

154

W.-T. Fu and V. Liao

(even though individually they are far from perfect) to facilitate quality control of online information. Designs should therefore encourage more efficient flows of user recommendations to fully harness the potential of crowdsourcing of quality control.

References 1. Bikhchandani, S., Hirshleifer, D., Welch, I.: A theory of fads, fashion, custom, and cultural change in informational cascades. Journal of Political Economy 100(5), 992–1026 (1992) 2. Brossard, D., Lewenstein, B., Bonney, R.: Scientific knowledge and attitude change: The impact of a citizen science project. International Journal of Science Education 27(9), 1099–1121 (2005) 3. Cheung, M., Luo, C., Sia, C., et al.: Credibility of electronic word-of-mouth: Informational and normative determinants of on-line consumer recommendations. Int. J. Electron. Commerce 13(4), 9–38 (2009) 4. Czaja, S.J., Charness, N., Fisk, A.D., et al.: Factors predicting the use of technology: Findings from the center for research and education on aging and technology enhancement (create). Psychology & Aging 21(2), 333–352 (2006) 5. Eysenbach, G.: Credibility of health information and digital media: New perspective and implications for youth. In: Metzger, M.J., Flanagin, A.J. (eds.) The John d. And Catherine t. Macarthur Foundation Series on Digital Media and Learning, pp. 123–154. MIT Press, Cambridge (2008) 6. Kempe, D., Kleinberg, J., Tardos, E.: Maximizing the spread of influence through a social network. In: Proceedings of the Ninth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Washington, D.C. (2003) 7. Liao, Q.V., Fu, W.-T.: Effects of cognitive aging on credibility assessments of online health information. In: Proceedings of the 28th Annual ACM Conference on Computerhuman Interaction (CHI), Atlanta, GA (2010) 8. Salganik, M.J., Dodds, P.S., Watts, D.J.: Experimental study of inequality and unpredictability in an artificial cultural market. Science 311(5762), 854–856 (2006) 9. Sunstein, C.R.: Infotopia: How many minds produce knowledge. Oxford University Press, Oxford (2006) 10. Watts, D.J.: A simple model of global cascades on random networks. Proceedings of the National Academy of Sciences of the United States of America 99(9), 5766–5771 (2002)

Consumer Search, Rationing Rules, and the Consequence for Competition Christopher S. Ruebeck Department of Economics, Lafayette College Easton PA, USA 18042 [email protected] http://sites.lafayette.edu/ruebeckc/

Abstract. Firms’ conjectures about demand are consequential in oligopoly games. Through agent-based modeling of consumers’ search for products, we can study the rationing of demand between capacity-constrained firms offering homogeneous products and explore the robustness of analytically solvable models’ results. After algorithmically formalizing short-run search behavior rather than assuming a long-run average, this study predicts stronger competition in a two-stage capacity-price game. Keywords: contingent demand, agent-based modeling, random arrival.

1

Introduction

The residual demand curve, or contingent demand curve, refers to the firm’s conjectured demand given its assumptions about consumers and other firms’ behaviors. Rather than discussing firms’ beliefs about each other’s actions, this paper investigates our more basic assumptions about customers’ arrivals when firms are capacity-constrained. That is, it focuses on the allocation of customers between firms, the “rationing rule.” The literature has shown that duopoly outcomes hinge on these assumptions, and we will see that previous characterizations of aggregate consumer behavior may have unrecognized ramifications. When Bertrand [3] critiqued the Cournot [5] equilibrium in quantities, he asserted that two firms are enough to create the efficient market outcome: if the strategic variable is price rather than quantity, firms will undercut each other until price equals marginal cost. To instead reinforce the Cournot “quantity” equilibrium as abstracting from a “capacity” decision, Kreps and Scheinkman [7] showed that a two-stage game can support price competition with a capacity choice that is the same as the one-stage Cournot quantities. This paper’s discussion follows Davidson and Deneckere’s [6] demonstration that the Kreps and Scheinkman result depends on the assumption that rationing is efficient. They characterized efficient rationing as one extreme, with Beckmann’s [2] implementation of Shubik’s [11] alternative, proportional rationing, occupying the 

Funded in part through grants HSD BCS-079458 and CPATH-T 0722211/0722203 from NSF. David Stifel, Ed Gamber, and Robert Masson provided helpful comments.

J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, pp. 155–162, 2011. c Springer-Verlag Berlin Heidelberg 2011 

156

C.S. Ruebeck

other extreme—although we will see behavior below that is outside those proposed extremes. Their analytic and simulation results cover rationing rules in general, and show that only the efficient rationing rule can provide the Kreps and Scheinkman result, arguing in addition that firms will not choose the efficient rationing rule if the two-stage game is extended to three stages that include the choice of rationing rule. The theory discussed thus far was developed for homogeneous goods, and it has been expanded to include asymmetric costs (Lepore [8]) and differentiated goods (Boccard and Wauthy [4]), but the results below are not about differentiation or variation in costs. Instead, I model search in which the consumer will stop once a low enough price is found at a firm with available capacity. Although this abstracts from the reasons for consumers’ need to search, doing so provides a baseline for the future both in terms of algorithms and results.

2

Rationing Rules

With two firms producing homogeneous goods in a simultaneous stage game, we study Firm 2’s best response to Firm 1’s decision of price and capacity. Both firms may be capacity constrained, but our discussion focuses on Firm 2’s unconstrained price choice as it considers Firm 1’s constrained decision. Because Firm 1 cannot satisfy the entire market at its price p1 , Firm 2 is able to reach some customers with a higher price p2 > p1 . The rationing rule specifying which customers buy from each firm will be a stochastic outcome in the simulations to follow, but has not been treated that way in theoretical work thus far. The shape of the efficient rationing rule, shown in Figure 1A, is familiar from exercises ‘shifting the demand curve’ that begin in Principles of Economics. This rationing rule’s implicit assumption about allocating higher-paying customers to the capacity-constrained firm has been known at least since Shubik’s discussion of the efficient rationing rule [12] [11]. The demand curve D(p) shifts left by the lower-priced firm’s capacity k1 , as depicted by the figure’s darker lines. Recognizing that Firm 2 can also be capacity-constrained, residual demand when p2 > p1 is D(p2 |p1 ) = min(k2 , max(0, D(p2 ) − k1 )) . (1) The rationing of customers between the two firms is indicated by the figure’s lightly dashed lines, showing that those units of the good with highest marginal benefit are not available to the higher-priced firm. Although the efficient rationing rule would appear to be an attractive characterization of a residual demand curve “in the long run,” firms are typically making decisions and conjectures about each other’s prices in the short run. The long run outcome occurs as a result of firms’ shorter-term decisions, their longrun constraints, their anticipations of each other’s decisions, and the evolution of the market, including both consumers’ and firms’ reactions to the shorter-term decisions. Game theoretic characterizations of strategic behavior in the short run are important in determining the long run, as the market’s players not only react to each other but also anticipate each other’s actions and reactions. Yet,

157

price

price

Consumer Search and Competition

MB(k1)

D(p)

D(p) D(p2|p1)

D(p2|p1) p1

p1

k1 k1

D(p1)

quantity

(A) The efficient rationing rule

k1

D(p1)

quantity

(B) The proportional rationing rule

Fig. 1. Standard rationing rules D(p2 |p1 ) when Firm 1 chooses capacity k1 and prices at p1 in a homogeneous goods market described by demand curve D(p)

in assuming the efficient rationing rule, the level of abstraction may penetrate further in some dimensions of the problem than others, leading to an unattractive mismatch between the short run one-shot game and long run outcomes. After discussing proportional rationing’s residual demand curve, the discussion and simulations below will show that even the short run abstraction may be mismatched with the descriptive reasoning behind the rationing rule. Both the proportional residual demand curve depicted in Figure 1B and the efficient demand curve in Figure 1A match market demand D(p) for p2 < p1 : they both take Bertrand’s assumption that all customers buy from the lowerpriced firm if they can. The simulation results below will not have the same characteristic. In the case of a “tie”, p2 = p1 , the literature makes two possible assumptions on the discontinuity. Davidson and Deneckere assume a discontinuity on both the right and the left of (a.k.a. above and below) p1 with the two firms splitting demand evenly, up to their production capacities. Allen and Hellwig [1] assume a discontinuity only to the left (below) p1 , pricing at equality by extending the upper part of the proportional rationing curve. It may be more accurate to characterize the upper part of the proportional rationing curve as an extension of the outcome when prices are equal, a point to which I will return below. The key feature of the proportional rationing rule is that all buyers with sufficient willingness to pay have a chance of arriving at both firms. That is, now all consumers with willingness to pay at or above p2 > p1 may buy from Firm 2. Thus in Figure 1B the market demand curve is “rotated” around its vertical intercept (the price equal to the most any customer will pay) rather than “shifting” in parallel as in Figure 1A. Describing the realization of this contingent demand curve between p1 and the maximum willingness to pay is not as intuitive as it is for the efficient contingent demand curve, although the highest point is easy: the maximum of of D(p) must also be the maximum of D(p2 |p1 ) because both firms have a chance at every customer.

158

C.S. Ruebeck

To justify the D(p2 |p1 ) proportional demand curve at other prices p2 > p1, the literature describes some specifics on consumers’ arrival processes and then connects these two points with a straight line, so that “the residual demand at any price is proportional to the overall market demand at that price” (Allen and Hellwig), or as Beckmann first states it, “When selling prices of both duopolists are equal, total demand is a linear function of price. When prices of the two sellers differ, buyers will try as far as possible to buy from the low-price seller. Those who fail to do so will be considered a random sample of all demanders willing to buy at the lower price.” The proportion adopted draws directly from the efficient rationing rule: the fraction of demand remaining for Firm 2 if Firm 1 receives demand for its entire capacity (up to Firm 2’s capacity and only greater than zero if market demand is greater than Firm 1’s capacity). Formally, the proportional contingent demand for Firm 2 when p2 > p1 is    D(p1 ) − k1 D(p2 |p1 ) = min k2 , max 0, D(p2 ) . (2) D(p1 ) 1 )−k1 The key feature of this rationing rule is the constant fraction D(p that D(p1 ) weights the portion of market demand Firm 2 receives at price p2 . First, this fraction is the demand firm 2 receives when prices are equal, and second this fraction of market demand D(p2 ) remains constant as Firm 2 increases price. The shape of the residual demand curve below will show that neither of these features need be true as we consider consumers’ arrival.

3

Implementing Random Arrival

We now turn to implementing consumer arrival and search rules. For the efficient rationing rule, consumers would be sorted according to the discussion of Figure 1A. They would not arrive in random order: all customers with willingnessto-pay (WTP) greater than the marginal benefit at k1 , MB(k1 ), are sent to the capacity constrained, lower-price firm. Somehow these customers rather than others get the good, perhaps by resale (Perry [9]), but it is not the point of this investigation to model such a process. Instead, we will have consumers that arrive in random order—one of the assumptions of the conventional proportional rationing rule, but just part of it. The other part is that the customers that arrive first are always allocated to the lower-priced firm. Relaxing this second part of the assumption is what drives the new results below. 3.1

Random Arrival, Random First Firm

Buyers arrive in random order at two capacity-constrained sellers. As a buyer arrives he chooses randomly between sellers; if the first seller’s price is higher than the buyer’s WTP, the buyer moves on to the other one, or does not buy at all if both prices are higher than his WTP. As each later buyer arrives, she can only consider sellers that have not yet reached their capacity constraints.

Consumer Search and Competition

159

As in the analytic models described above, we take the perspective of Firm 2 given Firm 1’s chosen price and capacity. Labeling one firm as “1” does not indicate that buyers in the simulation arrive first at either location; each buyer has an equal chance of arriving first at either Firm 1 or Firm 2. Figure 2A presents an example of the simulation results. Each of the lines in the plot (some overlapping each other) shows a statistic of 1300 runs; the statistics are min, max, mean, and median quantity for Firm 2 at each price p2 . The plot overlays six 1300-run sets, capturing the stochastic residual demand curve facing Firm 2 with no capacity constraint while the other firm sets price p1 = 9 and capacity k1 = 7. Market demand is q = 24 − p. Each of the 23 buyers has unitary demand at price D23 = 23, D22 = 22, . . . , D1 = 1. The proportional rationing rule is also shown; It is piecewise-linear, connecting the outcomes for p = 0, 8, 9, and 24. (The line segment connecting p = 8 and 9 reflects the discrete nature of demand in the simulation.) The efficient rationing rule is not shown, but it would maintain the slope of the lower line from p2 = p1 = 9 all the way to the vertical axis. The curved line in the plot is the mean residual demand received by Firm 2. As foreshadowed above, it is to the right not only of the efficient rationing rule, but the proportional rationing rule as well. The median is the series of stair steps around that line.

p2 = 10 mean q2 = 7.4 median q2 = 7

(B) p2 = 6 mean q2 = 10.7 median q2 = 10

(A)

(C)

Fig. 2. Simulation results. Firm 2’s residual demand curve with no capacity constraint and market demand q = 23 − p. Firm 1 has price p1 = 9 and capacity k1 = 7.

Consider first the plot’s horizontal extremes. Across the six 1300-run sets, Firm 2’s contingent demand has multiple maxima and minima q2 (horizontal axis) for some p2 (vertical axis) values. Because the 1300 simulated runs are not always enough to establish the true maxima and minima for all customer arrival outcomes, the distribution’s tail is (i) thinner in the minima for higher p2 in the upper left of the plot and (ii) thinner in the maxima for lower p2 in the lower

160

C.S. Ruebeck

right of the plot. Why? (i) The minimum q2 at any p2 is 7 units to the left of the market demand curve (but not less than 0), reflecting the capacity constraint of Firm 1. Thus the minimum contingent demand for Firm 2 is less likely to occur when its price p2 is high because it is less likely that all high-value buyers arrive first at Firm 2 when there are few possible buyers. (ii) The theoretical maximum at each price p2 is the entire demand curve: by chance all the customers may arrive first at Firm 2. It is more difficult to reach that maximum in a 1300-run sample at lower prices p2 , when there are more potential customers. Returning to the central part of the distribution, the mean and median lines, we see that—unlike the assumptions behind Figure 1B—the residual demand curve’s distribution is not always centered on the interval between its minimum and maximum. Figure 2B and Figure 2C illustrate the skewed distributions for two of Firm 2’s prices, p2 = 6 and 10. With these overall differences in hand, now consider, in the three areas of the graph, the differences in these simulations as compared to the theory pictured in Figure 1. In the lower part of the curve, the median is actually equal to the minimum for all prices p2 < p1 . The mean is pulled away from the median by the upper tail of the skewed distribution. The explanation for the observed behavior in the extremes applies here. Note, too, in the lower part of the curve that demand is smaller than the usual homogeneous goods assumption: The lower-priced firm does not capture the entire market. This is a consequence of maintaining the random arrival assumption for all prices in the simulation, not just for p2 > p1 as usually assumed (following Bertrand) in the analytical work described above. At equal prices p2 = p1 , the outcome is continuous in p2 , a point to consider again below. Finally, we saw that in the upper part of the curve p2 > p1 , the mean residual demand curve no longer changes linearly with price: the two firms share the market with varying proportions rather than the usual constant proportion assumption. Note that the traditional proportional residual demand curve (Davidson and Deneckere) provides Firm 2 with a “better” result than that provided by the efficient residual demand curve because it provides a larger share of the market. The results in Figure 2 are even better for the higher-priced firm! Firm 2’s proportion of the market grows as p2 increases, rather than remaining a constant proportion, because Firm 1’s capacity constraint improves the chance that more of the higher-value customers may happen to arrive first at Firm 1 instead of Firm 2. 3.2

Random Arrival, to the Lowest Price If Constraint Unmet

Next consider a specification of behavior that lies midway between the proportional rationing rule and the results in Figure 2. In this case, consumers will continue to search if they do not receive a large enough consumer surplus, WTP − p. This specification could be viewed as redundant because we have already specified both that consumers have a willingness-to-pay and that they do not know or have any priors on the firms’ prices. It may make more sense to instead specify those priors or updating of priors; I will address this further in the closing section.

Consumer Search and Competition

161

In Figure 3, the residual demand curve changes smoothly from the varyingproporition in Figure 2 to the proportional demand curve. When p2 > p1 , the choice of some consumers to continue searching pulls the demand curve back towards the proportional prediction. When p2 < p1 , the choice to continue searching pushes the demand curve out towards the usual homogeneous goods assumption. At p2 = p1 , we have a discontinuity for any search threshold that is not zero (the case of Figure 2).

Threshold changes from 0.1 to 0.2

Threshold changes from 0.2 to 0.3

Threshold changes from 0.4 to 0.5

Fig. 3. Continue search when consumer surplus is not a large enough fraction of WTP

4

Conclusions and Further Work

Allen and Hellwig state that consumers go to firms on a “first-come first-served basis” and that, “Firm j with the higher price pj meets the residual demand of those consumers—if any—who were unable to buy at the lower price.” The results above focused directly on a first-come first-served rule, with the result that some consumers with pi < WTP < pj may instead arrive too late to be served by the lower-priced firm and thus not buy. The resulting contingent demand curve provides a varying proportion of market demand at price pj . This approach recognizes that the searching consumers may know very little about the prices available before they arrive at the firms, and this assumption is unusual in search models. The limited literature that considers this type of search includes Rothschild [10] and Telser [13], but their analyses still require rather strong assumptions to arrive at a model amenable to analytic methods. Telser’s investigation finishes with the conclusion (emphasis added), “If the searcher is ignorant of the distribution, then acceptance of the first choice drawn at random from the distribution confers a lower average cost than more sophisticated procedures for a wide range of distributions. In most cases these experiments show that it simply does not pay to discover and patronize the lower price sellers . . . [and] we face the problem of explaining how a seller wishing to stress lower prices than his rivals can attract customers.” Thus, one direction to continue this work is to investigate the effect of this demand curve on equilibrium outcomes, moving from a one-shot static analysis to a dynamic interaction. Davidson and Deneckere show that any contingent

162

C.S. Ruebeck

demand curve different from the efficient demand curve must provide more market power (a larger quantity effect) to the higher-priced firm when the other firm is capacity-constrained. Thus firms in the first stage have greater incentive to avoid being capacity-constrained and choose capacities greater than the Cournot quantity, and leading to an equilibrium price support in the second-stage game that is closer to the Bertrand outcome of marginal cost pricing. The results of these simulations show that the competitive pressure is even stronger when we take seriously the assumptions underlying uninformed consumer serach. Setting a price above one’s competitor leads to less loss of demand, so that the market power that remains for the higher-priced firm is even larger than in the case of proportional rationing. At the same time, undercutting is less profitable: the lower-priced unconstrained firm does not receive all demand as assumed generally in the literature on homogeneous products. Thus we can expect that further analysis in a dynamic setting will provide outcomes that are even more competitive than found in existing analytical models.

References 1. Allen, B., Hellwig, M.: Bertrand-Edgeworth duopoly with proportional residual demand. International Economic Review 34(1), 39–60 (1993) 2. Beckmann, M.: Edgeworth-Bertrand duopoly revistied. In: Henn, R. (ed.) Operations Research Verfahren III, Meisenhein. Verlag Anton Hain, Meisenhein (1965) 3. Bertrand, J.: Th´eorie math´ematique de la richesse sociale (review). Journal des Savants 48, 499–508 (1883) 4. Boccard, N., Wauthy, X.: Equilibrium vertical differentiation in a Bertrand model with capacity precommitment. International Journal of Industrial Organization 28(3), 288–297 (2010) 5. Cournot, A.: Researches into the Mathematical Principles of Wealth. Macmillan, London (1838) 6. Davidson, C., Deneckere, R.: Long-run competition in capacity, short-run competition in price, and the Cournot model. The RAND Journal of Economics 17(3), 404–415 (1986) 7. Kreps, D., Scheinkman, J.: Precommitment and bertrand competition yield Cournot outcomes. The Bell Journal of Economics 14(2), 326–337 (1983) 8. Lepore, J.: Consumer rationing and the Cournot outcome. B.E. Journal of Theoretical Economics, Topics 9(1), Article 28 (2009) 9. Perry, M.: Sustainable positive profit multiple-price strategies in contestable markets. Journal of Economic Theory 32(2), 246–265 (1984) 10. Rothschild, M.: Searching for the lowest price when the distribution of prices is unknown. The Journal of Political Economy 82(4), 689–711 (1974) 11. Shubik, M.: A comparison of treatments of a duopoly problem. Econometrica 23(4), 417–431 (1955) 12. Shubik, M.: Strategy and Market Structure: Competition, Oligopoly, and the Theory of Games. John Wiley & Sons, New York (1959) 13. Telser, L.: Searching for the lowest price. American Economic Review 63(2), 40–49 (1973)

Pattern Analysis in Social Networks with Dynamic Connections Yu Wu1 and Yu Zhang2 1

Department of Computer Science Standford University [email protected] 2 Department of Computer Science Trinity University [email protected]

Abstract. In this paper, we explore how decentralized local interactions of autonomous agents in a network relate to collective behaviors. Most existing work in this area models social network in which agent relations are fixed; instead, we focus on dynamic social networks where agents can rationally adjust their neighborhoods based on their individual interests. We propose a new connection evaluation rule called the Highest Weighted Reward (HWR) rule, with which agents dynamically choose their neighbors in order to maximize their own utilities based on the rewards from previous interactions. Our experiments show that in the 2-action pure coordination game, our system will stabilize to a clustering state where all relationships in the network are rewarded with the optimal payoff. Our experiments also reveal additional interesting patterns in the network. Keywords: Social Network, Dynamic Network, Pattern Analysis.

1 Introduction Along with the advancement of modern technology, computer simulation is becoming a more and more important tool in today's researches. The simulation of large scale experiments which originally may take people months or even years can now be run within minutes. Computer simulation has not only saved researchers a great amount of time but also enabled them to study many macro topics that were impossible to study experimentally in the past. In this situation, Multi-Agent System (MAS) has emerged as a new area that facilitates the study of large scale social networks. Researchers have invented many classic networks, such as random networks, scalefree networks (Albert and Barbási 2002), small world networks (Watts 1999), to list a few. Although these networks have successfully modeled many social structures, they all share the weakness of being static. In today's world, many important virtual networks, such as e-commerce networks, social network services, etc., have much less stable connections among agents and thus the network structures will constantly change. Undoubtedly, the classic networks will fail to capture the dynamics of these networks. Out of this concern, (Zhang and Leezer 2009) develops the network model HCR J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, pp. 163–171, 2011. © Springer-Verlag Berlin Heidelberg 2011

164

Y. Wu and Y. Zhang

(Highest Cumulative Reward) that enables agents to update their connections through a rational strategy. In this way, any classic network can be easily transformed into a dynamic network through the HCR rule. In their model, agents evaluate their neighbors based on their performance history, and all interactions in history are equally weighted. However, people may argue that the evaluation function is not very realistic since in the real world recent events may have a greater influence on people than long past ones. Motivated by this fact, we extend HCR to a new rule called HWR (Highest Weighted Reward). The difference is that HWR allows the agents to use a discount factor to devalue past interactions with their neighbors. The more recent the interaction is, the more heavily the corresponding reward is weighted in the evaluation function. In our research, we identified certain patterns from the simulation results and then demonstrated the existence of the pattern empirically.

2 Related Work To our view, the social simulation conducted on MAS can go into two directions: One is to build a model as realistic as possible for a specific scenario, such as the moving crowds in an airport. The other one is to build an abstract model that captures the essence of human activities and can be used to model many different social networks of the same kind. Our work here belongs to the latter category. Research in this area usually focuses on the study of social norms. One of the most basic questions that can be asked about the social norms is whether the agents in the social network will eventually agree on one solution. Will the whole network converge on one solution? The results vary drastically in different networks and under different algorithms. The agents can converge into various patterns and with different time period. Generally, there are two categories in this area studying the emergence of social norms. The first category studies social norms in static networks. The second studies the evolving or dynamic network. In the first category of static network, one of the most significant findings was by (Shoham and Tennenholtz 1997). They proposed the HCR (Highest Current Reward) rule. In each timestep, an agent adopting HCR rule switches to a new action if and only if the total rewards obtained from that action in the last time step are greater than the rewards obtained from the currently chosen action in the same time period. The authors simulated various networks under different parameter configurations (memory update frequency, memory restart frequency, and both) to experiment on an agent’s effectiveness in learning about their environment. They also theoretically proved that under HCR rule the network will always converge to one social norm. Another important work done in the first category is the Generalized Simple Majority (GSM) rule proposed by (Delgado 2002). Agents obeying the GSM rule will change to an alternative strategy if they have observed more instances of it on other agents than their present action. This rule generalizes the simple majority rule. As the randomness β→∞, the change of state will occur as soon as more than half of the agents are playing a different action. In that case, GSM will behave exactly as the simple majority rule does. The second category of research on agents’ social behavior is the study of evolutionary networks or dynamic networks, such as (Borenstein 2003, Zimmermann 2005). Here researchers investigate the possible ways in which an agent population

Pattern Analysis in Social Networks with Dynamic Connections

165

may converge onto a particular strategy. The strong assumptions usually make the experiments unrealistic. For example agents often have no control over their neighborhood. Also, agents do not employ rational selfish reward maximizing strategies but instead often imitate their neighbors. Lastly, agents are often able to see the actions and rewards of their neighbors, which is unrealistic in many social settings. In the second category of the research, (Zhang and Leezer 2009) proposed the Highest Rewarding Neighborhood (HRN) rule. The HRN rule allows agents to learn from the environment and compete in networks. Unlike the agents in (Borenstein 2003), which can observe their neighbors’ actions and imitate it, the HRN agents employ selfish reward maximizing decision making strategy and are able learn from the environment. Under the HRN rule, cooperative behavior emerges even though agents are selfish and attempt only to maximize their own utility. This arises because agents are able to break unrewarding relationships and therefore are able to maintain mutually beneficial relationships. This leads to a Pareto-optimum social convention, where all connections in the network are rewarding. This paper proposes the HWR rule, which is extended from the HRN rule. In HRN rule, the agent values its neighbors based on the all rewards collected from them in the past, and all interactions are weighted equally throughout the agent's history. However, this will cause the agent to focus too much on the history of the neighbor and fail to respond promptly to the neighbor’s latest action. Therefore, we introduce the HWR rule, which introduces a time discount factor that helps the agents to focus on the recent history.

3 Highest Weighted Reward (HWR) Neighbor Evaluation Rule 3.1 HWR Rule The HWR rule is based on a very simple phenomenon we experience in our everyday lives: human beings tend to value recent events more than events which happened a long time ago. To capture this feature, we introduce a time discount factor, which is usually smaller than 1, in order to linearly devalue the rewards collected from past interactions. When time discount factor equals 1, the HWR rule will behave in the same way as the HRN rule does. Notice that both HWR and HRN are neighbor evaluation rules, which means the purpose of these rules is to help the agents value their connections with neighbors more wisely. In other words, these rules can only make decisions regarding connection choosing, but will not influence the agent's action choosing decisions. According to the HWR rule, an agent will maintain a relationship if and only if the weighted average reward earned from that relationship is no less than a specified percentage of the weighted average reward earned from every relationship. Next we carefully go through one cycle of HWR evaluation step-by-step. 1.

For each neighbor, we have a variable named TotalReward to store the weighted total reward from that neighbor. In each turn, we update the TotalReward through the following equation in which c represents the discount factor: TotalReward = TotalReward × C + RewardInThisTurn

(1)

166

Y. Wu and Y. Zhang

2. In a similar way, we keep a variable named GeneralTotalReward to store the weighted total reward the agent got from all neighbors in the history. In each turn, we update the GeneralTotalReward through the following equation: GeneralTotalReward = GeneralTotalReward × C + TotalRewardInThisTurn / NumOfNeighbors (2) Here TotalRewardInThisTurn needs to be divided by the NumOfNeighbors in each turn in order to find the average reward per connection. Since we want to calculate the average reward of all interactions, the total reward in each round needs to be divided by the number of interactions in that turn. 3. When we start to choose non-rewarding neighbors, we first calculate the average reward for every agent. The AvgReward is calculated in the following way: suppose the agent has been playing with the neighbor for n turns, then:

AvgReward =

TotalReward TotalReward = 2 n −1 1 + c + c + ... + c 1 − c n−1 1− c

(3)

4. Calculate the average total reward. Similar to step 3:

GeneralTotalReward GeneralTotalReward = (4) 1 + c + c 2 + ... + c n−1 1 − c n−1 1− c 5. Compare the ratio of AvgReward/AvgTotalReward with the threshold. If the former is greater than the later, then we keep the neighbor; if not, we regard that agent as a bad neighbor. In this way, the agent can evaluate its neighbor with an emphasis on recent history. AvgTotalReward =

The above explanation has assumed an ideal environment in which agents have infinity amount of memory for ease of exposition. In reality and our experiments, agents have a limited amount of memory and can only remember the interactions of a certain number of turns. 3.2 Discount Factor Range Generally speaking, the HWR rule can work with different games and in various networks. Since this rule just adds a dynamic factor into the network, it does not contradict other aspects of the network. However, in some special cases, a certain experiment setup may make some discount factor value meaningless. Assume we are playing a pure coordination game (Guillermo and Kuperman 2001) with payoff table (C: Cooperate; D: Defect): Table 1. Two-Person Two-Action Pure Coordination Game

C D

C 1 -1

D -1 1

Pattern Analysis in Social Networks with Dynamic Connections

167

Also assume the agents only want to keep the neighbors whose average reward is positive for them. Then in such a case, no matter how many rounds an agent has been playing with a neighbor, its most recent reward will dominate the whole reward sequence. To be more specific, if an agent receives a reward of 1 from a neighbor in the last interaction but has constantly received -1 in the previous n trials, the cumulative reward will be 1−c−c2−…− cn. If c is too low, the cumulative reward will still be bigger than 0. We can find the range of c through the following derivation: 1−c−c2−…−cn > 0 ⇒ 1 > c+c2+…+cn ⇒ 1 > c(1+cn)/(1−c)

(5)

When n→∞, cn→0, then

⇒1 > c/(1−c) ⇒ c < 0.5

(6)

Therefore, we usually limit c in range of 0.5 to 1. In fact, for most cases c will take a value much higher than 0.5. Since the past rewards were devalued exponentially, a small c value will cause the past rewards to be practically ignored very soon. In our common experiment setup, we let the agent have a memory size of 30. In this case, if c=0.5, then we see the oldest action will be devalued by a factor of 0.529≈1.86E-9, which makes the reward totally insignificant. In fact, even for a relatively high c value such as 0.9, the last factor will still be as small as 0.929≈0.047. Therefore, just in case the past rewards will be devalued too heavily, we usually keep the c in the range of 0.8 to 1. 3.3 Pattern Predictions Based on the HWR rule, we make some predictions about the outcome of the network pattern. Argument: Given a pure coordination game, placing no constraints on the initial choices of action by all agents, and assuming that all agents employ the HWR rule, then the following holds: • For every ε > 0 there exists a bounded number M such that if the system runs for M iterations then the probability that a stable clustering state will be reached is greater than 1 − ε. • If a stable clustering state is reached then the all agents are guaranteed to receive optimal payoff from all connections. In order to further clarify the argument, the following definitions are given: Stable Clustering State: A network reaches a stable clustering state when all the agents in the network belong to one and only one closed cluster. Closed Cluster: A set of agents forms a closed cluster when there does not exist any connection between any agent in the set and another agent outside of the set, and also all the agents in the set are playing the same action. Basically, what the argument claims is that the whole network will eventually converge into one single cluster of same action or two clusters with different actions, and once in such a state, the rewards collected from the network will be maximized.

168

Y. Wu and Y. Zhang

4 Experiment The operating procedure can be shown in the flow chart. Agents play with each other

Update actions based on the current neighborhood

Evaluate each connection based on the rewards just received

Break up with bad neighbors and update neighborhood

All experiments are conducted in an environment that has the following properties: 1. All trials are run in a random network having 1000 to 10000 agents. 2. Agents play with each other the pure coordination game defined in Table 1. 3. The number of connections in the network remains the same throughout the trial. 4. Every time a connection is broken, both agents have a 50% chance to gain the right to connect to a new neighbor. Exactly one of the two agents will find a new neighbor. This restriction guarantees the number of connections remains the same. 5. All agents have a limited memory size. 6. All agents adopt obey the simple majority rule, which means the agent will always choose the action that the majority of its neighbors used in the last turn. If there is the same number of neighbors adopting different actions, the agent will not change its current action. 4.1 Two-Cluster Stable Clustering State First, we show the emergence of the stable clustering state. Since the final outcome can have two different patterns with either one or two final stable clusters, we present the patterns in two parts. First, let's check the situation where the network eventually converges into one single cluster. The following experiments are conducted with the parameters: Number of Runs: 50, Time Steps: 30, Network Size: 5000, Neighborhood Size: 10, Reward Discount: 0.9, Threshold: 0.9.

Fig. 1. Network Converges into Two Clusters

Pattern Analysis in Social Networks with Dynamic Connections

169

Fig. 1 shows the average results collected from 50 trials. We see that at the end of each trial, all agents have split into two camps and entered into a unique stable clustering state. First, we notice that the lines that represent the numbers of C and D agents almost merge together in the middle of the plot. Since the pay-off matrix for C and D actions is symmetric and there is no difference between these two actions except different symbolic names, the average numbers of C and D agents should be both around 2500, just as the plot shows. Here in order to capture the dynamics of the network, we introduce a new term named perfect agent. An agent will be counted as a perfect agent only when all of his neighbors plays the same action as he does. In Figure 1, while the numbers of C and D agents remain almost constant, the number of perfect agents rises steadily from 0 to 5000 throughout the trial. The increasing number of perfect agents shows the network is evolving from a chaotic state towards a stable clustering state. At the same time, we see the number of broken connections is steadily dropping towards 0. As there are more and more agents becoming perfect agents, connections between neighbors are broken less and less frequently. 4.2 One-Cluster Stable Clustering State Now we present the experiment results where the agents all adopt the same action at last and merge into one cluster. In order to show the results more clearly, we have run 50 trials and selected 23 trials where C eventually becomes the dominant action. The experiments are carried out with the following parameters: Number of Runs: 50, Time Steps: 30, Network Size: 5000, Neighborhood Size: 100, Reward Discount: 0.9, Threshold: 0.9. In Fig. 2, we see number of C agents is steadily rising while number of D agents dropping until the whole network converges into a single cluster. The number of perfect agents and broken connections behaves in a manner similar to that in Fig. 1.

Fig. 2. Network Converges into One Single Cluster

4.3 Pattern Possibilities with Various Network Sizes The above experiments demonstrate that the network sometimes converges into one single cluster and sometimes into two clusters. After a series of simulations, we found

170

Y. Wu and Y. Zhang

out that the size of neighborhood has a significant influence over the probability that a network will converge into one cluster or not. In order to show this, we created 16 networks with different sizes and ran 50 trials with each network. Here are the parameters: Number of Runs: 50, Time Steps: 30, Network Size: 1000, Reward Discount: 0.9, Threshold: 0.9. Fig. 3 shows the number of one-cluster trials and twocluster trials with each network.

Fig. 3. Network Converges into One Single Cluster

5 Conclusion The HWR rule enables agents to increase their rewards not only through switching actions but also through updating neighborhoods. As the experiment results show, agents adopting the HWR rule are able to change the network structure to maximize their own interests. Even though the agents behave selfishly, the network still reaches a Pareto-optimum social convention at last.

Acknowledgement This work was supported in part by the U.S. National Science Foundation under Grants IIS 0755405 and CNS 0821585.

References 1. Abramson, G., Kuperman, M.: Social Games in a Social Network. Physical Review (2001) 2. Albert, R., Barbási, A.L.: Statistical Mechanics of Complex Networks. Modern Physics, 47–97 (2002) 3. Shoham, Y., Tennenholtz, M.: On the Emergence of Social Conventions: Modeling, Analysis and Simulations. Artificial Intelligence, 139–166 (1997) 4. Delgado, J.: Emergence of Social Conventions in Complex Networks. Artificial Intelligence, 171–175 (2002)

Pattern Analysis in Social Networks with Dynamic Connections

171

5. Borenstein, E., Ruppin, E.: Enhancing Autonomous Agents Evolution with Learning by Imitation. Journal of Artificial Intelligence and Simulation of Behavior 1(4), 335–348 (2003) 6. Zimmermann, M., Eguiluz, V.: Cooperation, Social Networks and the Emergence of Leadership in a Prisoners Dilemma with Adaptive Local Interactions. Physical Review (2005) 7. Watts, D.J.: Small Worlds. Princeton University Press, Princeton (1999) 8. Zhang, Y., Leezer, J.: Emergence of Social Norms in Complex Networks. In: Symposium on Social Computing Applications (SCA 2009), The 2009 IEEE International Conference on Social Computing (SocialCom 2009), Vancouver, Canada, August 29-31, pp. 549–555 (2009)

Formation of Common Investment Networks by Project Establishment between Agents Jes´ us Emeterio Navarro-Barrientos School of Mathematical and Statistical Sciences Arizona State University, Tempe, AZ 85287-1804, USA

Abstract. We present an investment model integrated with trust and reputation mechanisms where agents interact with each other to establish investment projects. We investigate the establishment of investment projects, the influence of the interaction between agents in the evolution of the distribution of wealth as well as the formation of common investment networks and some of their properties. Simulation results show that the wealth distribution presents a power law in its tail. Also, it is shown that the trust and reputation mechanism proposed leads to the establishment of networks among agents, presenting some of the typical characteristics of real-life networks like a high clustering coefficient and short average path length. Keywords: complex network, trust-reputation, wealth distribution.

1

Introduction

Recently, different socio-economical problems have been modeled and investigated using agent-based simulations, this approach presents more flexible, realistic and simple conceptual modeling perspectives. Many important contributions to this field are provided by the research group called agent-based computational economics (ACE) [9]. Different ACE models have been proposed to study for example the loyalty from buyers to sellers [2], investors and brokers in financial markets [3], and to understand the emergence of networks between agents [9], among others. The topology of the networks emerging from the interaction between agents is usually analyzed using statistical methods [1]. The main goal of this paper is to improve the understanding of two main components in ACE models: (i) the economical component that describes the dynamics of the wealth distribution among agents; and (ii) the social component that describes the dynamics of loyalty, trust and reputation among agents. For this, we integrate a wealth distribution model where agents invest a constant proportion of their wealth, and a network formation model where agents interact with each other to establish investment projects. J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, pp. 172–179, 2011. c Springer-Verlag Berlin Heidelberg 2011 

Formation of Common Investment Networks

2 2.1

173

The Model Wealth Dynamics

Consider an agent-based system populated with N agents, where each agent posses a budget xk (t) (measure of its “wealth” or “liquidity”) that evolves over time given the following dynamic:   xk (t + 1) = xk (t) 1 + rm (t) qk (t) + a(t), (1) where rm (t) denotes the return on investment (RoI) that the agent k receives from its investment qk (t) in project m, qk (t) denotes a proportion of investment, i.e the fraction or ratio of the budget of agent k that the agent prefers to invest in a market and a(t) denotes an external income, which for simplicity we assume to be constant, a(t) = 0.5. In this model, agent k invests a portion qk (t)xk (t) of its total budget at every time step t, yielding a gain or loss in the market m, expressed by rm (t). Similar wealth models have been presented in [4,5,7] where the dynamics of the investment model are investigated using some results from the theory of multiplicative stochastic processes [8]. Note that this approach assumes that the market, which acts as an environment for the agent, is not influenced by its investments, i.e. the returns are exogenous and the influence of the market on the agent is simply treated as random. This is a crucial assumption which makes this approach different from other attempts to model real market dynamics [3]. 2.2

Trust-Reputation Mechanisms and Project Establishment

In order to launch a particular investment project m at time t, a certain minimum amount of money Ithr needs to be collected among the agents. The existence of the investment threshold Ithr is included to enforce the interaction between agents, as they need to collaborate until the following condition is reached: Im (t) =

Nm 

qk (t) xk (t) ≥ Ithr ,

(2)

k

where Nm is the number of agents collaborating in the particular investment project m. For simplicity, it is assumed that each agent participates in only one investment project at a time. The first essential feature to be noticed for the formation of common investment networks is the establishment of preferences between agents. It is assumed that the decision of an agent to collaborate in a project will mainly depend on the previous history it has gained with other agents. Consider an agent k which accepts to collaborate in the investment project m initiated by agent j. Thus, agent k receives the following payoff at time t: pkj (t) = xk (t) qk (t) rm (t).

(3)

174

J.-E. Navarro-Barrientos

Reiterated interactions between agent k and agent j lead to different payoffs over time that are saved in a trust weight: wkj (t + 1) = pkj (t) + wkj (t) e−γ ,

(4)

where γ represents the memory of the agent. The payoffs obtained from previous time steps t may have resulted from the collaborative action of different agents, however, these are unknown to agent k, i.e agent k only keeps track of the experience with the initiator of the project, agent j. Furthermore, in order to mirror reality, it is assumed that from the population of N agents only a small number J are initiators, i.e. J  N , where the reputation of an initiator j can be calculated as follows (for more on trust and reputation models see [6]): Wj (t + 1) =

N 

wkj (t).

(5)

k=0

At every time step t an initiator is chosen randomly from the population and assigned with an investment project. The initiator randomly tries to convince other agents to invest in the project until an amount larger than the threshold Ithr has been collected. For this, we use a Boltzmann distribution to determine the probability that the contacted agent k may accept the offer of agent j: eβwkj (t) τkj (t) = J , βwki (t) i=1 e

(6)

where in terms of the trust weight wkj , the probability τkj (t) considers the good or bad previous experience with agent j with respect to the experience obtained with other initiators; and β denotes the greediness of the agent, i.e. how much importance does the agent give to the trust weight wkj . In order to make a decision, agent k uses a technique analogous to a roulette wheel where each slice is proportional in size to the probability value τkj (t). Thus, agent k draws a random number in the interval (0, 1) and accepts to invest in the project of agent j if the segment of agent j in the roulette spans the random number. Finally, an initiator j stops to contact other agents if either the investment project has reached the threshold Ithr or if all agents in the population have been asked for collaboration. Now, if the project could be launched it has to be evaluated. The evaluation should in general involve certain “economic” criteria that also reflects the nature of the project. However, for simplicity we assume that the failure or success of an investment project m is randomly drawn from a uniform distribution, i.e. rm (t) ∼ U (−1, 1).

3

Results of Computer Simulations

We performed some simulations using the parameter values in Table 1, for simplicity, we assume that the initial budget xk (0) is the same for all agents and the

Formation of Common Investment Networks

175

Table 1. Parameter values of the computer experiments for the model of formation of common investment networks Parameter Num. of agents Num. of initiators Num. of time steps Investment threshold

Value N = 104 J = 100 t = 105 Ithr = 9

Parameter Initial budget Initial trust Memory Greediness

Value xk (0) = 1 wkj (0) = 0 γk = 0.1 βk = 1

proportion of investment is assumed to be constant and the same for all agents, i.e. qk (t) = q = const. Fig. 1 (left) shows the evolution of the budget distribution over time for q = 0.5. Note that the probability distribution of the budget converges to a stationary distribution with a power law in the tail, a known property of investment models based on multiplicative processes repelled from zero [8,5]. Fig. 1 (right) shows the distribution of the budget at time step t = 105 for different constant proportions of investment q. Note that even for a large number of time steps, the budget distribution for agents with a proportion of investment of q = 0.1 has not yet converged to a stationary distribution, whereas for q = 0.5 and q = 0.9, the distribution has reached a stationary state. We examine the evolution over time of the budget and reputation of the initiators to understand their role in the dynamics of the investment model. For the sake of clarity, the rank-size distribution of the budget of the initiators is shown in Fig. 2 (left), note that richer initiators would grow ever richer. It is also interesting to examine the evolution of the budget of the initiator with the largest and the smallest budget at the end of the simulation. This is shown in the inset of Fig. 2 (left), note that the budget of the best agent was not always increasing over time, which shows the relevance of stochastic returns in the dynamics. Fig. 2 (right) shows the rank-size distribution of the reputation of the initiators, Eq. (5), note that the distribution does not change over time, meaning that only for a small number of agents there is a significant increase or decrease on reputation over time. Moreover, it can be shown that the average value of the reputation has a shift to larger positive values over time. This occurs due to an aggregation of external incomes a(t), Eq. (1), into the dynamics of the trust weights, Eq. (4). Moreover, the inset in Fig. 2 (right) shows the reputation of the best and the worst initiator indicating the presence of no symmetrical positive/negative reputation values. The influence of other parameters in the dynamics of the model was also analyzed, however, for brevity, we discuss only the role of the number of initiators J in the dynamics. It can be shown that, if the number of initiators is small, then more investors will be willing to invest in their projects. This leads to a larger amount of investment that can be collected by the initiators. It was shown in Fig. 1 that the tail of the wealth distribution has a power law distribution, it can be shown that the larger the number of initiators J, the larger the slope

176

J.-E. Navarro-Barrientos

of the power law. The reason for this is that a small number of initiators collect more money from the investors leading to larger profits and looses which over time lead to wider distributions than for a large number of initiators.

0

0

10

10

q=0.1 q=0.5 q=0.9

t=10000 t=30000 t=50000 t=100000

-2

probability

probability

-2

10

-4

10

-6

10

-4

10

-6

10

10 0

1

10

2

10

3

10

4

10

x

0

10

10

10

1

2

3

10

4

10

x

10

Fig. 1. Probability distribution of the budget: (left) evolution over time for a constant proportion of investment q = 0.5; (right) for different proportions of investment q after t = 105 time steps. Additional parameters as in Table 1. 3

10

10

0.004

3

best 2

0.003

0.005

t=10000 t=50000 t=100000

0.004 0.003

xj

10

10

1

worst

0.002

0

20000

40000

60000

t

80000

100000

0.001 0 -0.002

0.001

1

best

0.002

worst

-0.001

Wj

2

10

Wj

xj

10

20000

40000

60000

80000

100000

t

10

0

t=10000 t=50000 t=100000

0

10 1

10

20

30

40

50

Rank

60

70

80

90

100

-0.001 1

10

90

100

Rank

Fig. 2. For initiators, evolution over time of the rank-size distribution of (left) budget and (right) reputation. Insets show the best and the worst budget and reputation, respectively. Additional parameters as in Table 1.

4

Structure of Common Investment Networks

In this section, we analyze the topology of the networks. For simplicity and brevity, we focus our analysis on the influence of different constant proportions of investment in the emerging structure of the common investment networks. For this, we run different computer experiments for a small population of agents N = 103 (N is also the number of nodes in the network) and other parameter values as in Table 1. The first experiment investigates the influence of the proportion of investment in the properties of the network. Fig. 3 shows the networks emerging from the common investment model for different proportion of investment q at time step t = 103 . Note that these networks have two types of nodes, the black nodes represent investors and the gray nodes represent initiators. Based on visual

Formation of Common Investment Networks

177

impression, the density of the network decreases with respect to the proportion of investment. This occurs because agents investing more also tend to loose more, which leads to more mistrust. However, from the visual representation of the network it is not possible to draw many conclusions from the dynamics of the networks. Thus, we need to rely on other methods to characterize the networks, typically the following statistical properties of the networks are used: number of links V , maximal degree kmax (the degree of the highest degree vertex of the network), average path length l (the average shortest distance between any pair of nodes in the network), and clustering coefficient C (measures the transitivity of the network).

Fig. 3. Common investment networks for different proportions of investment at time step t = 103 : (left) q = 0.1, (right) q = 0.5 and (bottom) q = 0.9. A link between agents represents a positive decision weight, i.e wkj > 0. For N = 103 investors (black nodes) and J = 10 initiators (gray nodes). Additional parameters as in Table 1.

On the following we present the most important properties of the networks emerging from our model. First, it can be shown that both the number of links

178

J.-E. Navarro-Barrientos

V as well as the maximal degree kmax of the network increase over time proportionally to a power-law function, where the slope of the power law decreases if the proportion of investment increases. This basically says that the density of the networks increase over time and can be characterized. Moreover, it can be shown that the clustering coefficient C is large for small proportions of investment. This occurs because a small proportion of investment leads to a higher clustering in the network due to the mistrust that large losses generate in the investors. Table 2 shows some of the most important characteristics for different number of investors N and initiators J for a large number of time steps, i.e. t = 105 . For each network we indicate the average degree k (the first moment of the degree distribution), the average path length l and the clustering coefficient C. For comparison reasons we include the average path length lrand = log (N )/ log (k) and the clustering coefficient Crand = k /N that can be obtained from a random network with the same average degree k of the investment networks. Note that the average degree k increases with respect to the system size. In general, the networks show a small average path length l ≈ 2, meaning that any investor or initiator in the network is on average connected to each other by two links. The clustering coefficient of the investment networks is larger than the clustering coefficient of a random network, this indicates the presence of transitivity in our networks. This occurs mainly because of the large number of investors connected to initiators. Finally, note that the clustering coefficient of the networks decreases with respect to N , this is in qualitative agreement with properties of small-world networks [1]. Table 2. Properties of the networks for different number of investors N and initiators J. The properties measured are: V the number of links, kmax the maximal degree, k the average degree, l the average path length and C the clustering coefficient. For proportion of investment q = 0.5, t = 105 and additional parameters as in Table 1. N 1000 2000 3000 10000

5

J 10 20 30 100

V 4847 19972 41073 134279

kmax 517 1050 1475 1477

k 0.9694 3.9944 8.2146 26.86

l 2.05766 1.99365 1.99314 2.15630

C 0.74557 0.71337 0.71130 0.24136

lrand 5.488 3.8018 2.7989

Crand 0.0009694 0.0019972 0.0027382 0.0026860

Conclusions and Further Work

This article presents the initial investigations on a model for formation of common investment networks between agents. The model integrates two models, an investment model where agents place investments proportional to their budget, and a trust and reputation model where agents keep track of their experience with initiators of projects, establishing a common investment network. The most important conclusions that can be drawn from the model here presented are the following: (i) the budget of the agents reaches a stationary distribution after some time steps and presents a power law distribution on the tail, a property

Formation of Common Investment Networks

179

discussed in other investment models [5,7,8]; and (ii) the topology of the investment networks emerging from the model present some of the typical characteristics of real-life networks like a high clustering coefficient and short average path length. We focused our investigations on the feedback that describes the establishment and reinforcement of relations among investors and initiators. This is considered a “social component” of the interaction between agents and it was shown how this feedback process based on positive or negative experience may lead to the establishment of networks among agents. Furthermore, we noted that the external income sources play an important role on the dynamics, leading to nonsymmetrical distributions of trust and reputation among agents. We note also that further experiments are needed for different memory γ and greediness β values to understand their influence in the dynamics of the networks.

Acknowledgements We thank Prof. Frank Schweitzer for his advice during these investigations and Dr. Ramon Xulvi-Brunet for the program Xarxa to analyze the networks.

References 1. Albert, R., Barab´ asi, A.-L.: Statistical mechanics of complex networks. Rev. Mod. Phys. 74, 47–97 (2002) 2. Kirman, A., Vriend, N.J.: Evolving market structure: An ACE model of price dispersion and loyalty. J. Econ. Dyn. Control 25, 459–502 (2001) 3. LeBaron, B.: Agent-based computational finance: Suggested readings and early research. J. Econ. Dyn. Control 24, 679–702 (2000) 4. Navarro, J.E., Schweitzer, F.: The investors game: A model for coalition formation. In: Czaja, L. (ed.) Proceedings of the Workshop on Concurrency, Specification & Programming, (CS & P 2003), vol. 2, pp. 369–381. Warsaw University, Czarna (2003) 5. Navarro-Barrientos, J.E., Cantero-Alvarez, R., Rodrigues, J.F.M., Schweitzer, F.: Investments in random environments. Physica A 387(8-9), 2035–2046 (2008) 6. Sabater, J., Sierra, C.: Review on computational trust and reputation models. Artificial Intelligence Review 24(1), 33–60 (2005) 7. Solomon, S., Richmond, P.: Power laws of wealth, market order volumes and market returns. Physica A 299(1-2), 188–197 (2001) 8. Sornette, D., Cont, R.: Convergent multiplicative processes repelled from zero: Power laws and truncated power laws. Journal of Physics 1(7), 431–444 (1997) 9. Tesfatsion, L., Judd, K. (eds.): Handbook of Computational Economics, vol. 2. Elsevier, Amsterdam (2006)

Constructing Social Networks from Unstructured Group Dialog in Virtual Worlds Fahad Shah and Gita Sukthankar Department of EECS University of Central Florida 4000 Central Florida Blvd, Orlando, FL [email protected], [email protected]

Abstract. Virtual worlds and massively multi-player online games are rich sources of information about large-scale teams and groups, offering the tantalizing possibility of harvesting data about group formation, social networks, and network evolution. However these environments lack many of the cues that facilitate natural language processing in other conversational settings and different types of social media. Public chat data often features players who speak simultaneously, use jargon and emoticons, and only erratically adhere to conversational norms. In this paper, we present techniques for inferring the existence of social links from unstructured conversational data collected from groups of participants in the Second Life virtual world. We present an algorithm for addressing this problem, Shallow Semantic Temporal Overlap (SSTO), that combines temporal and language information to create directional links between participants, and a second approach that relies on temporal overlap alone to create undirected links between participants. Relying on temporal overlap is noisy, resulting in a low precision and networks with many extraneous links. In this paper, we demonstrate that we can ameliorate this problem by using network modularity optimization to perform community detection in the noisy networks and severing cross-community links. Although using the content of the communications still results in the best performance, community detection is effective as a noise reduction technique for eliminating the extra links created by temporal overlap alone. Keywords: Network text analysis, Network modularity, Semantic dialog analysis.

1 Introduction Massively multi-player online games and virtual environments provide new outlets for human social interaction that are significantly different from both face-to-face interactions and non-physically-embodied social networking tools such as Facebook and Twitter. We aim to study group dynamics in these virtual worlds by collecting and analyzing public conversational patterns of Second Life users. Second Life (SL) is a massively multi-player online environment that allows users to construct and inhabit their own 3D world. In Second Life, users control avatars, through which they are able to explore different environments and interact with other avatars in a variety of ways. One of the most commonly used methods of interaction in Second J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, pp. 180–187, 2011. c Springer-Verlag Berlin Heidelberg 2011 

Constructing Social Networks from Unstructured Group Dialog in Virtual Worlds

181

Life is basic text chat. Users are able to chat with other users directly through private instant messages (IMs) or to broadcast chat messages to all avatars within a given radius of their avatar using a public chat channel. Second Life is a unique test bed for research studies, allowing scientists to study a broad range of human behaviors. Several studies on user interaction in virtual environments have been conducted in SL including studies on conversation [1] and virtual agents [2]. The physical environment in Second Life is laid out in a 2D arrangement, known as the SLGrid. The SLGrid is comprised of many regions, with each region hosted on its own server and offering a fully featured 3D environment shaped by the user population. The current number of SL users is estimated to be 16 million, with a weekly user login activity reported in the vicinity of 0.5 million [3]. Although Second Live provides us with rich opportunities to observe the public behavior of large groups of users, it is difficult to interpret who the users are communicating to and what they are trying to say from public chat data. Network text analysis systems such as Automap [4] that incorporate linguistic analysis techniques such as stemming, named-entity recognition, and n-gram identification are not effective on this data since many of the linguistic preprocessing steps are defeated by the slang and rapid topic shifts of the Second Life users. This is a hard problem even for human observers, and it was impossible for us to unambiguously identify the target for many of the utterances in our dataset. In this paper, we present an algorithm for addressing this problem, Shallow Semantic Temporal Overlap (SSTO), that combines temporal and language information to infer the existence of directional links between participants. One of the problems is that using temporal overlap as a cue for detecting links can produce extraneous links and low precision. To reduce these extraneous links, we propose the use of community detection. Optimizing network modularity reduces the number of extraneous links generated by overly generous temporal co-occurrence assumption but does not significantly improve the performance of SSTO. There has been previous work on constructing social networks of MMORPG players, e.g., [5] looks at using concepts from social network analysis and data mining to identify tasks. Recent work [6] has compared the relative utility of different types of features at predicting friendship links in social networks; in this study we only examine conversational data and do not include information about other types of Second Life events (e.g., item exchanges) in our social networks.

2 Method 2.1 Dataset We obtained conversation data from eight different regions in Second Life over fourteen days of data collection; the reader is referred to [7] for details. To study user dialogs, we examined daily and hourly data for five randomly selected days in the eight regions. In total, the dataset contains 523 hours of information over the five days (80,000 utterances) considered for the analysis across all regions. We did a hand-annotation of one hour of data from each of the regions to serve as a basis for comparison. While there are corpora like [8], there has not been any body of work with an online chat corpus in a multi-user, open-ended setting — the characteristics of this dataset. In

182

F. Shah and G. Sukthankar

such situations it is imperative to identify conversational connections before proceeding to higher level analysis like topic modeling, which is itself a challenging problem. We considered several approaches to analyzing our dialog dataset, ranging from statistical NLP approaches using classifiers to corpus-based approaches using tagger/parsers; however we discovered that there is no corpus available for group-based online chat in an open-ended dialog setting. It is challenging even for human labelers to annotate the conversations themselves due to the large size of the dataset and the ambiguity in a multi-user open-ended setting. Furthermore, the variability of the utterances and the nuances such as emoticons, abbreviations and the presence of emphasizers in spellings (e.g., “Yayyy”) makes it difficult to train appropriate classifiers. Parser/tagger-based approaches perform poorly due to the lack of corpus and inclusion of non-English vocabulary words. Consequently, we decided to investigate approaches that utilize non-linguistic cues such as temporal co-occurrence. Although temporal co-occurrence can create a large number of false links, many aspects of the network group structure are preserved. Hence we opted to implement a two-pass approach: 1) create a noisy network based solely on temporal co-occurrence, 2) perform modularity detection on the network to detect communities of users, and 3) attempt to filter extraneous links using the results of the community detection. 2.2 Modularity Optimization In prior work, community membership has been successfully used to identify latent dimensions in social networks [9] using techniques such as eigenvector-based modularity optimization [10] which allows for both complete and partial memberships. As described in [10], modularity (denoted by Q below) measures the chances of seeing a node in the network versus its occurrence being completely random; it can be defined ki kj as the sum of the random chance Aij − 2m summed over all pairs of vertices i, j that fall in the same group, where si equals 1 if the two vertices fall in the same group and -1 otherwise: 1  ki kj Q= (Aij − )si sj . (1) 4m 2m k k

i j If B is defined as the modularity matrix given by Aij − 2m , which is a real symmetric matrix and s column vectors whose elements are s then Equation 1 can be written as i n 1 T 2 Q = 4m to the eigeni=1 (uI s) βi , where βi is the eigenvalue of B corresponding n vector u (ui are the normalized eigenvectors of B so that s = i ai ui and ai = uTi s). We use the leading eigenvector approach to spectral optimization of modularity as described in [11] for the strict community partitioning (s being 1 or -1 and not continuous). For the maximum positive eigenvalue we set s = 1 for the corresponding element of the eigenvector if its positive and negative otherwise. Finally we repeatedly partition a group of size ng in two and  calculate the change in modularity measure given by 1 Δq = 4m i,jg [Bij − δij kg Bik ]si sj , where δij is the Kronecker δ symbol, terminating if the change is not positive and otherwise choosing the sign of s (the partition) in the same way as described earlier.

Constructing Social Networks from Unstructured Group Dialog in Virtual Worlds

183

2.3 Shallow Semantics and Temporal Overlap Algorithm (SSTO) Because of an inability to use statistical machine learning approaches due to the lack of sufficiently labeled data and absence of a tagger/parser that can interpret chat dialog data, we developed a rule-based algorithm that relies on shallow semantic analysis of linguistic cues that commonly occur in chat data including mentions of named entities as well as the temporal co-occurrence of utterances to generate a to/from labeling for the chat dialogs with directed links between users. Our algorithm employs the following types of rules: salutations: Salutations are frequent and can be identified using keywords such as “hi”, “hello”, “hey”. The initial speaker is marked as the from user and users that respond within a designated temporal window are labeled as to users. questions: Question words (e.g., “who”, “what”, “how”) are treated in the same way as salutations. We apply the same logic to requests for help (which are often marked by words such as “can”, “would”). usernames: When a dialog begins or ends with all or part of a username (observed during the analysis period), the username is marked as to, and the speaker marked as from. second person pronouns: If the dialog begins with a second person pronoun (i.e., “you”, “your”), then the previous speaker is considered as the from user and the current speaker the to user; explicit mentions of a username override this. temporal co-occurrences: Our system includes rules for linking users based on temporal co-occurrence of utterances. These rules are triggered by a running conversation of 8–12 utterances. This straightforward algorithm is able to capture sufficient information from the dialogs and is comparable in performance to SSTO with community information, as discussed below. 2.4 Temporal Overlap Algorithm The temporal overlap algorithm consists of using the temporal co-occurrence to construct the links. It exploits the default timeout in Second Life (20 minutes) and performs a lookup for 20 minutes beginning from the occurrence of a given username and constructs an undirected link between the speakers and this user. This process is repeated for all users within that time window (one hour or day) in 20 minute periods. This algorithm gives a candidate pool of initial links between the users without considering any semantic information. Later, we show that incorporating community information from any source (similar time overlap or SSTO based) and on any scale (daily or hourly) enables us to effectively prune links, showing the efficacy of mining community membership information. 2.5 Incorporating Community Membership Our dataset consists of 5 randomly-chosen days of data logs. We separate the daily logs into hourly partitions, based on the belief that an hour is a reasonable duration for social interactions in a virtual world. The hourly partitioned data for each day is used to

184

F. Shah and G. Sukthankar

generate user graph adjacency matrices using the two algorithms described earlier. The adjacency matrix is then used to generate the spectral partitions for the communities in the graph, which are then used to back annotate the tables containing the to/from labeling (in the case of the SSTO algorithm). These annotations serve as an additional cue capturing community membership. Not all the matrices are decomposable into smaller communities so we treat such graphs of users as a single community. There are several options for using the community information — we can use the community information on an hourly- or daily basis, using the initial run from either SSTO or the temporal overlap algorithms. The daily data is a long-term view that focuses on the stable network of users while the hourly labeling is a fine-grained view that can enable the study of how the social communities evolve over time. The SSTO algorithm gives us a conservative set of directed links between users while the temporal overlap algorithm provides a more inclusive hypothesis of users connected by undirected links. For the SSTO algorithm, we consider several variants of using the community information: SSTO: Raw SSTO without community information; SSTO+LC: SSTO (with loose community information) relies on community information from the previous run only when we fail to make a link using language cues. SSTO+SC: SSTO (with strict community information) always uses language cues in conjunction with the community information. For the temporal overlap algorithms, we use the community information from the previous run. TO: Raw temporal overlap algorithm without community information; TO+DT Temporal overlap plus daily community information; TO+HT Temporal overlap plus hourly community information.

3 Results In this section we summarize the results from a comparison of the social networks constructed from the different algorithms. While comparing networks for similarity is a difficult problem [12], we restrict our attention to comparing networks as a whole in terms of the link difference (using Frobenius norm) and a one-to-one comparison for the to and from labelings for each dialog on the ground-truthed subset (using precision and recall). 3.1 Network Comparison Using the Frobenius Norm We constructed a gold-standard subset of the data by hand-annotating the to/from fields for a randomly-selected hour from each of the Second Life regions. It is to be noted that there were instances where even a human was unable to determine the person addressed due to the complex overlapping nature of the dialogs in group conversation in an open ended setting (Table 2).

Constructing Social Networks from Unstructured Group Dialog in Virtual Worlds

185

To compare the generated networks against this baseline, we use two approaches. First we compute a Frobenius norm [13] for the adjacency matrices from the corresponding networks. The Frobenius norm is the matrix norm of an M × N matrix A and is defined as:  M N   A =  |aij |2 . (2) i=1 j=1

The Frobenius norm directly measures whether the two networks have the same links and can be used since the networks consists of the same nodes (users). Thus, the norm serves as a measure of error (a perfect match would result in a norm of 0). Table 1 shows the results from this analysis. Table 1. Frobenius norm: comparison against hand-annotated subset SSTO SSTO+LC SSTO+SC

TO

TO+DT TO+HT 162.00 130.08 20.29 54.88 58.44 49.89 37.54 38.98 73.15 77.82 62.62 71.02 53.64 50.97 39.01 41.10

Help Island Public Help People Island Mauve Morris Kuula Pondi Beach Moose Beach Rezz Me

35.60 62.23 48.45 24.67 32.12 20.63 17.08 36.70

41.19 60.50 45.11 18.92 30.75 21.77 18.30 39.74

46.22 66.34 51.91 20.76 32.66 21.56 21.07 45.78

224.87 20.29 58.44 43.12 83.22 75.07 67.05 38.72

Total error

277.48

276.28

306.30

610.78 507.21 514.74

3.2 Direct Label Comparisons The second quantitative measure we present is the head-to-head comparison of the to/from labelings for the dialogs using any of the approaches described above (for SSTO) against the hand annotated dialogs. This gives us the true positives and false positives for the approaches and allows us to see which one is performing better on the dataset, and if there is an effect in different Second Life regions. Table 2 shows the results from this analysis.

(a) Hand labeled network

(b) SSTO labeled network

(c) TO labeled network

Fig. 1. Networks from different algorithms for one hour in the Help Island Public region

186

F. Shah and G. Sukthankar Table 2. Precision/Recall values for one-to-one labeling comparison Help Help Mauve Morris Kuula Pondi Moose Rezz Island People Beach Beach Me Public Island Total Dialogs

360

184

128

179

227

144

128

97

Hand Labeled

recall total

0.6278 0.9076 0.9453 0.6983 0.8370 0.6944 0.6797 0.8866 226 167 121 125 190 100 87 86

SSTO+SC

match precision recall F-Score total

61 59 49 43 63 27 12 23 0.2607 0.6629 0.6364 0.4216 0.4632 0.3971 0.2105 0.4600 0.2699 0.3533 0.4050 0.3440 0.3316 0.2700 0.1379 0.2674 0.2652 0.4609 0.4204 0.3789 0.3865 0.3214 0.1667 0.3382 234 89 77 102 136 68 57 50

SSTO+LC

match precision recall F-Score total

61 51 37 39 52 26 12 15 0.3005 0.6456 0.6607 0.4643 0.4561 0.4194 0.2667 0.4688 0.2699 0.3054 0.3058 0.3120 0.2737 0.2600 0.1379 0.1744 0.2844 0.4146 0.4181 0.3732 0.3421 0.3210 0.1818 0.2542 203 79 56 84 114 62 45 32

SSTO

match precision recall F-Score total

76 68 51 45 66 30 20 27 0.3065 0.7083 0.6145 0.4500 0.4748 0.4225 0.3077 0.4576 0.3363 0.4072 0.4215 0.3600 0.3474 0.3000 0.2299 0.3140 0.3207 0.5171 0.5000 0.3617 0.4012 0.3509 0.2299 0.3724 248 96 83 100 139 71 65 59

4 Conclusion For the temporal overlap algorithm (TO), the addition of the community information reduces the link noise, irrespective of the scale — be it hourly or daily. This is shown by the decreasing value of the Frobenius norm in all the cases as compared to the value obtained using temporal overlap algorithm alone. In general shallow semantic approach (SSTO) performs the best and is only improved slightly by the loose incorporation of community information. For the SSTO algorithm, the daily or hourly community partition also does not affect the improvement. Table 2 shows how the dialog labeling generated from various algorithms agrees with the ground truth notations produced by a human labeler. Since TO only produces undirected links, we do not include it in the comparison. Plain SSTO generally results in a better precision and recall than SSTO plus either strict or loose community labeling. This is further confirmed from the visualizations for one hour of data in a day (both chosen randomly from the dataset) for each of the ground-truth, SSTO and TO as shown in figure 1, where the network from SSTO more closely resembles the ground-truth as compared to the on from TO. The challenging nature of this dataset is evident in the overall low precision and recall scores, not only for the proposed algorithms but also for human labelers. We attribute this largely to the inherent ambiguity in the observed utterances. Among the

Constructing Social Networks from Unstructured Group Dialog in Virtual Worlds

187

techniques, SSTO performs best, confirming that leveraging semantics is more useful than merely observing temporal co occurrence. We observe that community information is not reliably informative for SSTO but does help TO, showing that link pruning through network structure is useful in the absence of semantic information.

Acknowledgments This research was funded by AFOSR YIP award FA9550-09-1-0525.

References 1. Weitnauer, E., Thomas, N.M., Rabe, F., Kopp, S.: Intelligent agents living in social virtual environments – bringing Max into Second Life. In: Prendinger, H., Lester, J.C., Ishizuka, M. (eds.) IVA 2008. LNCS (LNAI), vol. 5208, pp. 552–553. Springer, Heidelberg (2008) 2. Bogdanovych, A., Simoff, S.J., Esteva, M.: Virtual institutions: Normative environments facilitating imitation learning in virtual agents. In: Prendinger, H., Lester, J.C., Ishizuka, M. (eds.) IVA 2008. LNCS (LNAI), vol. 5208, pp. 456–464. Springer, Heidelberg (2008) 3. Second Life: Second Life Economic Statistics (2009), http://secondlife.com/whatis/economy_stats.php (retrieved July 2009) 4. Carley, K., Columbus, D., DeReno, M., Bigrigg, M., Diesner, J., Kunkel, F.: Automap users guide 2009. Technical Report CMU-ISR-09-114, Carnegie Mellon University, School of Computer Science, Institute for Software Research (2009) 5. Shi, L., Huang, W.: Apply social network analysis and data mining to dynamic task synthesis to persistent MMORPG virtual world. In: Proceedings of Intelligent Virtual Agents (2004) 6. Kahanda, I., Neville, J.: Using transactional information to predict link strength in online social networks. In: Proceedings of the Third International Conference on Weblogs and Social Media (2009) 7. Shah, F., Usher, C., Sukthankar, G.: Modeling group dynamics in virtual worlds. In: Proceedings of the Fourth International Conference on Weblogs and Social Media (2010) 8. Mann, W.C.: The dialogue diversity corpus (2003), http://www-bcf.usc.edu/˜billmann/diversity/DDivers-site.htm (retrieved November 2010) 9. Tang, L., Liu, H.: Relational learning via latent social dimensions, in kdd 2009. In: Proceedings di of the 15th ACM SIGKDD International ti Conference on Knowledge, pp. 817–826. ACM, New York (2009) 10. Newman, M.: Modularity and community structure in networks. Proceedings of the National Academy of Sciences 103, 8577–8582 (2006) 11. Newman, M.: Finding community structure in networks using the eigenvectors of matrices. Phys. Rev. E 74, 036104 (2006) 12. Prulj, N.: Biological network comparison using graphlet degree distribution. Bioinformatics 23(2) (2007) 13. Golub, G.H., Loan, C.F.V.: Matrix Computations, 3rd edn. JHU Press (1996)

Effects of Opposition on the Diffusion of Complex Contagions in Social Networks: An Empirical Study Chris J. Kuhlman1 , V.S. Anil Kumar1 , Madhav V. Marathe1 , S.S. Ravi2 , and Daniel J. Rosenkrantz2 1

Virginia Bioinformatics Institute, Virginia Tech, Blacksburg, VA 24061, USA {ckuhlman,akumar,mmarathe}@vbi.vt.edu 2 Computer Science Department, University at Albany – SUNY, Albany, NY 12222, USA {ravi,djr}@cs.albany.edu

Abstract. We study the diffusion of complex contagions on social networks in the presence of nodes that oppose the diffusion process. Unlike simple contagions, where a person acquires information or adopts a new ideology based on the influence of a single neighbor, the propagation of complex contagions requires multiple interactions with different neighbors. Our empirical results point out how network structure and opposing perspectives can alter the widespread adoption of social behaviors that can be modeled as complex contagions. Keywords: social influence, opposition, complex contagions, networks.

1

Introduction and Motivation

Understanding how populations of interacting individuals change their views, beliefs, and actions has many applications. Among these are encouraging peaceful social changes, stopping or inciting mob behavior, spreading of information or ideologies, and sowing the seeds of distrust within a group [3, 5, 11]. Human behaviors in these situations are often complicated by significant risks. For example, people have been jailed for participating in peaceful demonstrations [4]. Also, in changing one’s ideology or acting against prevailing norms, one risks reprisals to oneself and family [10]. In these types of situations, people do not change their attitudes in a flippant manner. Granovetter [5] studied such phenomena and introduced the concepts of simple and complex contagions to describe how one’s social acquaintances can influence a person’s choices. A social network is a graph in which vertices represent people and an edge between two nodes indicates a certain relationship (e.g. the two corresponding persons are acquaintances of each other). A simple contagion is an entity such as a belief, information, rumor, or ideal that can propagate through a network such that a person acquires the contagion and is willing to pass it on based on a single interaction with one neighbor. A complex contagion, J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, pp. 188–196, 2011. c Springer-Verlag Berlin Heidelberg 2011 

Opposition to Complex Contagion Diffusion

189

in contrast, requires multiple interactions with (different) sources in order to accept the contagion and propagate it to others. Associated with these ideas is the concept of a threshold t, which is the number of neighbors with a particular view that is required for a person to change her mind to the same point of view. Thus, a simple contagion corresponds to t = 1 while a complex contagion corresponds to t > 1. Actions with greater consequences are generally modeled as complex contagions because one requires affirmation to change her beliefs [3]. Resistance offered by opponents is another complicating factor in the spread of a contagion. For example, if one seeks to change a person’s view, there will often be those with opposite opinions that seek to prevent the change. One approach to model this counter-influence is to treat these network nodes as having a fixed (old) opposing view that will not change, thus inhibiting diffusion. In this work, we carry out a simulation-based study of the diffusion of complex contagions on synthetic social networks. A diffusion instance (called an iteration) consists of the following. We specify a subset of nodes as initially possessing the contagion (corresponding to nodes in state 1), while all other nodes (which are in state 0) do not possess the contagion. We also specify a fraction f of failed nodes (following the terminology in [1]) that are in state 0 initially and will remain in state 0 throughout the diffusion process. Failed nodes are determined in one of two ways, following [1]: random failed nodes are chosen uniformly at random from the entire set of nodes; targeted failed nodes are nodes with the greatest degrees. These failed nodes represent the counter-influence that opposes the propagation. The diffusion process progresses in discrete time steps. At each time step, a node changes from state 0 to state 1 when at least t of its neighbors are in state 1, but nodes never change from state 1 to 0; hence, this is a ratcheted system [8, 9]. A diffusion instance ends when no state changes occur. We examine two types of networks. Scale-free (SF) networks possess hub nodes with high connectivity to other nodes, and are representative of many social networks [6]. Exponential decay (ED) networks have a tighter range of node degrees; their usefulness in modeling social behavior is discussed in [1]. Our empirical results provide insights for two perspectives: (1) how the diffusion process can spread widely even in the presence of failed nodes, and (2) how diffusion can be resisted by choosing the number and type of failed nodes. The significance of our results in modeling social behavior is discussed in Section 3, following discussions of the model and related work. Due to space limitations, a number of details regarding our experimental procedures and results are not included in this paper; these appear in a longer version [8].

2 2.1

Dynamical System Model and Related Work System Model and Associated Definitions

For studying diffusion in networks, we use the synchronous dynamical system (SyDS) model detailed in [8, 9]. Here we provide a brief account of it. In this model, each node of a social network represents an individual and each undirected edge {u, v} indicates that the contagion may propagate from u to v

190

C.J. Kuhlman et al.

or from v to u. Each node has a state value which may be 0 indicating that the node is unaffected (i.e., the contagion has not propagated to that node) or 1 indicating that the node has been affected (i.e., the contagion has propagated to that node). Once a node’s state changes to 1, it remains in that state for ever. State changes at the nodes of a social network are determined by local transition functions. Each of these is a Boolean t-threshold function which can be described as follows. The function fv at a node v has value 1 if the state of v is 1 or at least t of the neighbors of v are in state 1; else fv = 0. At each time step, all the nodes compute their local transition function values synchronously; the value of the function gives the new state of the node. We use the term tthreshold system to denote a SyDS in which the local transition function at each node is the t-threshold function. (The threshold t is the same for all the nodes.) When no node changes state, the system is said to have reached a fixed point. It is known that every such ratcheted SyDS reaches a fixed point within n steps, where n is the number of nodes in the underlying graph [8]. In the context of contagions, a fixed point marks the end of the diffusion process. For any SyDS, nodes which are in state 1 at time 0 are called seed nodes. The spread size is the number (or fraction) of nodes in state 1 (i.e., the number of affected nodes) at the end of the diffusion process. A cascade occurs when the spread size is a large fraction of all possible nodes; that is, most of the nodes that can be affected are indeed affected. An example to show that node failures can significantly affect the time needed to reach a cascade is presented in [8]. 2.2

Related Work

The average time required to reach a fixed point in ratcheted dynamical systems with relative thresholds (i.e., thresholds normalized by node degrees) was examined in [2, 3, 12]. The work presented in [2, 3] considered small-world and random networks generated by rewiring network edges. The focus was on measuring the average spread size as a function of time and the average number of time steps to cascade as a function of the rewiring probability. We know of no work on complex contagions that examines the time histories of individual diffusion instances to assess variations in behavior with respect to the time to cascade, with or without failed nodes. The effect of failed nodes on the ability of SF and ED graphs to propagate complex contagions using ratcheted dynamics was investigated by Centola [1]. His work used one average degree value dave = 4 and one threshold value (t = 2). We extend that work by examining comparable networks with twice the average degree (dave = 8) to examine the influence of this parameter. Other works have examined the thwarting of diffusion by specifying a set of blocking nodes tailored for a particular set of seed nodes (e.g., [9] and references therein). The focus here, in contrast, is network dynamics for blocking or failed nodes that are fixed a priori for all sets of seed nodes. Watts [12] primarily examines Erd˝os-R´enyi random graphs, which are different from ours, and his generating function approach cannot be extended directly to analyze the dynamics with targeted failed nodes.

Opposition to Complex Contagion Diffusion

3

191

Results and Their Significance

We now provide a summary of our results and their significance in modeling social behavior. (a) For the diffusion of desirable social changes that can be modeled as complex contagions (e.g. ideology), the required duration of commitment to achieve the goal can vary significantly depending on the sets of individuals that initiate the behavior and those who oppose it; in particular, the variation can be by a factor of 3 or more (Section 5.1). (b) Sustained resistance may be required to inhibit undesirable complex contagions. Vigilant monitoring may be necessary because a contagion may initially propagate very slowly, and only at a much later time reach a critical mass whereupon spreading is rapid. Furthermore, once the rate of diffusion increases, there is relatively little time for countermeasures (Section 5.1). (c) In critical applications where diffusion must be thwarted, the requisite number of opposing individuals can be much larger (e.g. by three orders of magnitude) compared to the number of individuals who initiated the behavior. Moreover, the opposing population may need to be 10% or more of all the nodes in the social network, particularly if the region from where a contagion may emerge is uncertain (Section 5.2). (d) Whether fostering or blocking diffusion, network structure matters. Swaying the view of a well-connected node in an SF network has a greater impact than one in an ED network – this effect is accentuated in graphs with higher average degree (Section 5.2). (e) Owing to the nature of ratcheted dynamical systems, it is instructive to separate the (1) probability of a cascade occurring and (2) the average cascade size, given that a cascade occurs. These single-parameter measures provide greater insight than computing a normalized average spread size (as done in [1]) over all diffusion instances (Section 5.3).

4

Networks and Experiments

One ED network and one SF network are evaluated in this study (these names were chosen to be consistent with [1]). Both are composed of 10000 nodes. The ED and SF networks have, respectively, 39402 and 39540 edges, average degrees of 7.88 and 7.91, and average clustering coefficients of 0.244 and 0.241. The rationale for choosing these networks is as follows. First, we wanted to investigate exponential decay and scale-free networks to be consistent with Centola’s work [1] so that we can compare the two sets of results. Centola used ED and SF networks with an average degree of 4. We chose networks with an average degree close to 8 for two reasons. First, we want to study the resilience properties of such networks when the average degree differs significantly from the threshold value. Second, there are social networks in the literature that have degree close to 8 [6], so that as part of a larger effort, we are planning to compare these synthetic networks to real social networks of the same average degree. We chose

192

C.J. Kuhlman et al.

an average clustering coefficient of about 0.24 to compare with the value 0.25 used in Centola’s work [1]. Network details are given in [8]. A diffusion instance, called an iteration, was described in Section 1. A simulation consists of 400 iterations, which is comprised of 20 failed node sets and 20 seed node sets for each failed node set. All iterations in a simulation utilize the same network, threshold, failed node determination method, and fraction f of failed nodes. Table 1 contains the parameters and their values used in this experimental study. Simulations were run with all combinations of the parameters in this table. Failed nodes are determined first, and then seed node sets are generated randomly from the remaining nodes. Further details regarding the choices of seed nodes and failed nodes are given in [1, 8]. Table 1. Parameters and values of parametric study Networks ED, SF

5

Failed Node Seeding Threshold, t Failed Node Methods Method Fraction, f random, targeted set one node and all 2 10−4 , 10−3 , 10−2 , neighbors to state 1 10−1 , 2 × 10−1

Experimental Results

This section presents the empirical results from our simulations on ED and SF networks. The results point out the impact of failures on the time needed to reach a cascade, the size of cascades and the probability of cascades. 5.1

Time to Reach a Cascade with a Fixed Number of Failures

Figures 1(a) and 1(b) depict the cumulative fraction of affected nodes as a function of time for the ED and SF networks, respectively, with random failures and f = 0.2. (Throughout this section, the captions for figures give the parameters used in generating the corresponding plots.) For each simulation, each of the 400 iterations is represented by a gray curve, and the average over the 400 diffusion instances is represented by the black curve. The ED results show significant scatter among iterations in the times for the diffusion to pick up speed (i.e., when the curves turn upwards) and for a fixed point to be reached (i.e., when the curves plateau). The time variation is roughly a factor of three. Note that some iterations have small spread sizes. It can be seen that the times to reach a fixed point are much shorter in Figure 1(b) for the SF network. The hub nodes of high degree, when not specified as failed nodes, greatly assist the spreading process. Essentially, hubs reduce the diameter of the SF network, compared to that of the ED network, so that if cascades occur, they occur quickly. The targeted failed node results for ED and f = 0.2 are qualitatively the same as those in Figure 1(a), but targeted failures in the SF network completely thwart all diffusion. Figure 1(c) shows that as f increases, variability in times to reach a fixed point increases for the ED network. From Figures 1(a) and 1(b), where the time

Opposition to Complex Contagion Diffusion

193

to reach a fixed point follows soon after diffusion speed up, the data in Figure 1(c) also reflect the variation in time to speed up as a function of f . If a goal is to spread an ideology or fad, Figures 1(a) and 1(b) also show that depending on seed and failed node sets and network structure, the required duration of commitment can vary by a factor of 10, with commensurate variations in costs.

(a)

(b)

(c)

Fig. 1. Cumulative fraction of affected nodes as a function of time for t = 2 and f = 0.2: (a) ED network and random failed nodes, and (b) SF network and random failed nodes. 400 iterations are shown in gray in each plot. In (c), the minimum, average, and maximum times to reach a fixed point for iterations with large spread sizes are plotted for the ED network and random failed nodes.

5.2

Effect of Failures on Cascade Size

In Figure 2 we combine our results with those of Centola [1] to examine the average normalized spread size for both types of networks as a function of random and targeted failures and average degree. The average normalized spread size S, i a measure used in [1, 3], is given by S = nk=1 Sk /ni where the sum is over the ni = 400 iterations and Sk is the spread size for the kth iteration. (Other measures are used later in this section.) Figure 2(a) shows results for the ED network. The dashed curves are for dave = 4, taken from Centola [1], while the solid curves are for dave = 8 and were generated in this work. For a given f value, the dave = 8 curves (solid lines) are above those for dave = 4. This is to be expected, since a greater average degree provides more neighbors from which a node can receive the contagion. The relationship between the magnitudes of dave and t is important. With dave = 4 and t = 2, a node v with degree only one less than the average needs every neighbor to acquire the contagion and pass it on (v requires two neighbors to receive the contagion and a third neighbor to which it can transmit). Hence, one would expect a significant difference when dave increases to 8. In this case, each node on average has 8 neighbors from which only 3 are needed to contract the contagion and help in passing it on. This is why the curves differ even with f as small as 10−5 , where there are no failed nodes: the network structure plays a role irrespective of failed nodes. We confirm these observations in Figure 2(b) for the SF network. The effect of a greater average degree on random failures (solid gray curve as compared to the dashed gray curve) is to drive the curve

194

C.J. Kuhlman et al.

significantly upward to the point that it is moving toward the shape of the solid gray curve for the ED network in Figure 2(a). These results for increasing dave are significant because real social and contact networks can have average degrees much larger, on the order of 20 or 50 or more [6, 7]. Further, targeted degree based failures have much more impact than random failures. Hence, from a practical standpoint, if there are relatively few people with a great number of neighbors (i.e., highly influential people), it is important to convert them to one’s point of view, for if they are in the opposition, they can markedly decrease S, and as we shall see in the next section, the probability of achieving widespread diffusion.

(a)

(b)

Fig. 2. The average normalized spread size S as a function of failed node fraction f for t = 2: (a) ED network and (b) SF network. Dashed and solid lines are for average degrees of 4 and 8, respectively. Dashed curves are taken from [1].

5.3

Effect of Failures on Cascade Probability

In discussing the previous results, we used the S measure defined in [1]. Here, we evaluate our simulation results in a different manner in order to gain greater insights into the network dynamics. Figure 3(a), generated for the SF network and targeted failed nodes, is constructed in the following fashion. In one simulation, say for f = 10−3 , 400 iterations are performed with 20 failed node sets and 20 seed sets. For each iteration, the fraction of affected nodes (at the fixed point) is obtained. These 400 values are arranged in increasing numerical order and are plotted as the black curve in Figure 3(a). The black curve lies along the horizontal axis from abscissa value of 0 until it rises sharply to 1.0 at an abscissa value of 0.45. The gray curve for f = 10−2 remains along the x-axis until it extends vertically at an abscissa value of 0.98. In this case, however, the curve turns horizontal at a final fraction of nodes affected of 0.9. The features of these curves are representative of the behavior of deterministic ratcheted SyDSs across a wide range of networks, both synthetic like those evaluated here and social networks generated from, for example, online data, and for diffusion processes with and without “blocking” nodes such as the failed nodes examined here (e.g., [8, 9]). Plots of this nature seem to be fundamental to ratcheted systems. The point of transition of a curve in Figure 3(a) is the probability of a cascade and the y-axis value after the transition is simply the cascade size for iterations that cascade. Although not apparent from Figure 3(a), this cascade size is

Opposition to Complex Contagion Diffusion

195

essentially constant. Other curves from this study and other types of studies [8, 9] also demonstrate this behavior. This means that we can plot the probability of a cascade and the average cascade size—each of which is a one-parameter characterization—as a function of f (one f value corresponding to each curve in Figure 3(a)). This is done in Figures 3(b) and 3(c). The black curves in these latter two figures were generated from the data in Figure 3(a) for targeted failures; the gray curves were generated from random failure data. If we compare these last two figures with the solid curves of Figure 2(b), it is obvious that the dominant driver in the change of S is the change in the probability of a cascade. As an example, consider f = 5 × 10−3 and targeted failures for dave = 8 in Figure 2(b), which gives an average normalized spread size of 0.3. If our goal is to prevent a cascade, it appears that this f is satisfactory. However, this conclusion is misleading because as Figure 3(c) illustrates, if widespread diffusion is achieved, then the contagion will reach over 90% of nodes.

(a)

(b)

(c)

Fig. 3. (a) For the SF network, ordered spread size for 400 iterations, for t = 2 for each curve. (b) Probability of a cascades as a function of f . (c) Average spread size for iterations that cascade as a function of f .

6

Conclusions and Future Work

Our results show that time to cascade for particular diffusion instances can vary significantly from mean behavior, that the effects of a threshold must be examined in light of network parameters such as degree distribution, and that for ratcheted systems, it is often possible to isolate the size of a cascade from the probability of a cascade. Future work includes study of realistic social networks. Acknowledgment. We thank the SBP2011 referees, our external collaborators, and members of NDSSL for their suggestions and comments. This work has been partially supported by NSF Nets Grant CNS- 0626964, NSF HSD Grant SES-0729441, NIH MIDAS project 2U01GM070694-7, NSF PetaApps Grant OCI-0904844, DTRA R&D Grant HDTRA1-0901-0017, DTRA CNIMS Grant HDTRA1-07-C-0113, DOE DE-SC0003957, US Naval Surface Warfare Center N00178-09-D-3017 DEL ORDER 13, NSF Netse CNS-1011769 and NSF SDCI OCI-1032677.

196

C.J. Kuhlman et al.

References 1. Centola, D.: Failure in Complex Social Networks. Journal of Mathematical Sociology 33, 64–68 (2009) 2. Centola, D., Eguiluz, V., Macy, M.: Cascade Dynamics of Complex Propagation. Physica A 374, 449–456 (2007) 3. Centola, D., Macy, M.: Complex Contagions and the Weakness of Long Ties. American Journal of Sociology 113(3), 702–734 (2007) 4. Euchner, C.: Nobody Turn Me Around. Beacon Press (2010) 5. Granovetter, M.: Threshold Models of Collective Behavior. American Journal of Sociology 83(6), 1420–1443 (1978) 6. Leskovec, J.: website, http://cs.stanford.edu/people/jure/ 7. Khan, M., Kumar, V., Marathe, M., Zhao, Z., Dutta, T.: Structural and Relational Properties of Social Contact Networks with Applications to Public Health Informatics, NDSSL Technical Report No. 09-066 (2009) 8. Kuhlman, C., Kumar, V., Marathe, M., Ravi, S., Rosenkrantz, D.: Effects of Opposition on the Diffusion of Complex Contagions in Social Networks, NDSSL Technical Report No. 10-154 (2010) 9. Kuhlman, C.J., Anil Kumar, V.S., Marathe, M.V., Ravi, S.S., Rosenkrantz, D.J.: Finding Critical Nodes for Inhibiting Diffusion of Complex Contagions in Social Networks. In: Balc´ azar, J.L., Bonchi, F., Gionis, A., Sebag, M. (eds.) ECML PKDD 2010. LNCS, vol. 6322, pp. 111–127. Springer, Heidelberg (2010) 10. Nafisi, A.: Reading Lolita in Tehran: A Memoir in Books. Random House (2003) 11. Nicholls, W.: Place, Networks, Space: Theorising the Geographies of Social Movements. Transactions of the Institute of British Geographers 34, 78–93 (2009) 12. Watts, D.: A Simple Model of Global Cascades on Random Networks. PNAS 99, 5766–5771 (2002)

Promoting Coordination for Disaster Relief – From Crowdsourcing to Coordination Huiji Gao, Xufei Wang, Geoffrey Barbier, and Huan Liu Computer Science and Engineering Arizona State University Tempe, AZ 85281 {Huiji.Gao,Xufei.Wang,Geoffrey.Barbier,Huan.Liu}@asu.edu

Abstract. The efficiency at which governments and non-governmental organizations (NGOs) are able to respond to a crisis and provide relief to victims has gained increased attention. This emphasis coincides with significant events such as tsunamis, hurricanes, earthquakes, and environmental disasters occuring during the last decade. Crowdsourcing applications such as Twitter, Ushahidi, and Sahana have proven useful for gathering information about a crisis yet have limited utility for response coordination. In this paper, we briefly describe the shortfalls of current crowdsourcing applications applied to disaster relief coordination and discuss one approach aimed at facilitating efficient collaborations amongst disparate organizations responding to a crisis. Keywords: Disaster Relief, Crisis Map, Crowdsourcing, Groupsourcing, Response Coordination, Relief Organization.

1

Introduction

Natural disasters have severe consequences including causalities, infrastructure and property damages, and civilian displacement. The 2004 Indian Ocean earthquake and tsunami killed 230,000 people in 14 countries. Hurricane Katrina occurred in 2005 killed at least 1,836 people and resulted in approximately $81 billion [4] in economical loss. The catastrophic magnitude 7.0 earthquake in Haiti on January 12, 2010 resulted in more than 230,000 deaths, 300,000 injuries, and one million homeless. Quality data collection from disaster scenes is a challenging and critical task. Timely and accurate data enables government and non-governmental organizations (NGOs) to respond appropriately. The popularity and accessibility of social media tools and services has provided a new source of data about disasters. For example, after the devastating Haiti earthquake in January 2010, numerous texts and photos were published via social media sites such as Twitter, Flickr, Facebook, and blogs. This is a type of data collection and information sharing that strongly leverages participatory social media services and tools known as crowdsourcing [3]. Crowdsourcing is characterized by collective contribution and has been adopted in various applications. For instance, Wikipedia depends J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, pp. 197–204, 2011. c Springer-Verlag Berlin Heidelberg 2011 

198

H. Gao et al.

on crowds to contribute wisdom without centralized management. Customer reviews are helpful in buying products from Amazon. Collaboration is a process to share information, resources, and cooperate among various organizations. Collaboration in disaster relief operations is usually decentralized because of the relative independence of relief organizations. Attempts have been made to help the relief community to enhance cooperation. Haiti Live 1 , which leverages web 2.0 technologies to collect data from various sources such as phones, the web, e-mail, Twitter, Facebook, etc., provides an up-to-date publicly available crisis map that is in turn available to relief organizations. However, without centralized control, it is difficult to avoid conflicts such as responding to the same request repeatedly by different organizations. An approach to enable efficient collaboration between disparate organizations during disaster relief operations is imperative in order for relief operations to be successful in meeting the needs of the people impacted by the crisis. This paper is organized as follows: we summarize related work in disaster relief management systems. We discuss the need for a collaboration system and highlight our approach.

2

Related Work

Dynes et al. [1] propose a disaster zone model which consists of five spatial zones of Impact, Fringe, Filter, Community, and Regional and eight socio-temporal disaster stages. Goolsby et al. [2] demonstrate the possibility in leveraging social media to generate community crisis maps and introduce an inter-agency map not only to allow organizations to share information, but also to collaborate, plan, and execute shared missions. The inter-agency map is designed to share (some) information between organizations if the organizations share the same platform or have similar data representation formats. In [5], Sophia and Leysia summarize 13 crisis-related mashups to derive some high-level design directions of next generation crisis support tools. Ushahidi is a crisis map platform created in 2007. The platform has been deployed in Kenya, Mexico, Afghanistan, Haiti, and other locations. Ushahidi can integrate data from various sources: phones, a web application, e-mail, and social media sites such as Twitter and Facebook. This platform uses the concept of crowdsourcing for social activism and public accountability to collectively contribute information, visualize incidents, and cooperate among various organizations. Sahana2 is an open source disaster management system which was started just after the 2004 Indian Ocean tsunami. It provides a collection of tools to help manage coordination and collaboration problems resulting from a disaster. The major functions are supporting the search for missing persons, coordinating relief efforts, matching donations to needs, tracking the status of shelters, and the reporting of timely information. Additionally, Sahana facilitates the management 1 2

http://haiti.ushahidi.com/ http://www.sahanafoundation.org/

Promoting Coordination for Disaster Relief

199

of volunteers by capturing their skills, availability, allocation, etc. This system has been deployed in several countries [6].

3

Why Crowdsourcing Falls Short for Disaster Relief

Although crowdsourcing applications can provide accurate and timely information about a crisis, there are three reasons why current crowdsourcing applications fall short in supporting disaster relief efforts. First, and most importantly, current applications do not provide a common mechanism specifically designed for collaboration and coordination between disparate relief organizations. For example, microblogs and crisis maps do not provide a mechanism for apportioning response resources. Second, current crowdsourcing applications do not have adequate security features for relief organizations and relief operations. For example, crowdsourcing applications that are publicly available for reporting are also publicly available for viewing. While this is important for providing information to the public it can create conflict when decision must be made about where and when relief resources are needed. Additionally, in some circumstances, relief workers themselves are targeted by nefarious groups. Publicizing the details of relief efforts can endanger relief workers. Third, data from crowdsourcing applications, while providing useful information, does not always provide all of the right information needed for disaster relief efforts. There are often duplicate reports and information essential for relief coordination is not readily available or easily accessible such as lists of relief resources or communication procedures and contact information for relief organizations.

4

How to Facilitate Disaster Relief

In order to facilitate efficient coordination during relief efforts, relief organizations need to leverage the information available from the group. Supplementing the crowdsourcing information that is available through social media, the relief organizations can contribute to a unified source of information customized for the group or “groupsourced.” We define groupsourcing as intelligently using information provided by a sanctioned group comprised of individuals with disparate resources, goals, and capabilities. Essentially, the response group is performing crowdsourcing but specialized for the entire group and taking it a few steps further. Using information provided from the crowd and the sanctioned members of the group, a relief response can be coordinated efficiently. In order to be most efficient and most effective in helping to resolve the crisis, the members of the response group should subscribe to the centralized administrative control of an information management system to ensure data integrity, data security, accuracy, and authentication will be realized for each member of the response group. One member of the response group can field the response coordination system ensuring equal access to other members of the response group. This data management system can leverage the latest social media technology and supplement functionality customized for the group.

200

5

H. Gao et al.

What an Application Needs

A disaster relief information system designed for better collaboration and coordination during a crisis should contain four technical modules: request collection, response, coordination, and statistics, as shown in Figure 1. User requests are collected via crowdsourcing and groupsourcing and stored in a requests pool after preprocessing. The system visualizes the requests pool on its crisis map. Organizations respond to requests and coordinate with each other through the crisis map directly. A statistics module runs in the background to help track relief progress. ACT (ASU Coordination Tracker) is an open disaster relief coordination system that implements the groupsourcing system architecture.

Fig. 1. Groupsourcing System Architecture

5.1

Request Collection

Relief organizations need two types of requests: requests from crowds (crowdsourcing) and requests from groups (groupsourcing). Crowdsourcing refers to requests submitted by people (e.g. victims, volunteers) who are not from certified organizations. Crowdsourcing data from disaster scenes are valuable for damage assessment and initial decision making. However, these data are usually subjective, noisy, and can be inaccurate. The groupsourcing requests originate from responding organizations such as United Nation, Red Cross, etc. The key difference between these two types of requests is the level of trustworthiness. Requests from the relief organizations are more likely objective and accurate compared to those from crowds. Sample requests are shown in Figures 2 and 3. Figure 4 shows an example crisis map (left picture) viewing fuel requests. Each green node represents an individual fuel request and each red node represents a request cluster, which consists of multiple individual requests (or additional request clusters) that are geographically close to each other. The number and the node size of each individual request represents the requested quantity, for example, 113 (left picture) means this request asks for 113 k-gallon fuels in that region. Individual requests in each cluster can be further explored by zooming-in (right picture).

Promoting Coordination for Disaster Relief

Fig. 2. Crowdsourcing

201

Fig. 3. Groupsourcing

Fig. 4. ACT Crisis Map

5.2

Request Response

A key aspect of efficient coordination is enabling relief organizations to respond in an organized and collaborative manner. Response actions need to be supported by the information system such that all of the relief organizations can contribute, receive, and understand the response options and planned response actions. The system must be available on heterogeneous hardware and software platforms (i.e. web services and contemporary commercially available portable devices). Figure 4 shows how a requests pool can be visualized on a crisis map. As the map resolution changes, nodes indicating the same type of requests will merge or split based on their location closeness. Here is how a response transaction may flow. Once an organization decides to respond to a request, it selects the node and adds it into a relief package. Responders are allowed to add multiple nodes of various categories into the relief package. The requests that are being fulfilled will be removed from the relief

202

H. Gao et al.

Fig. 5. ACT Response Process

Fig. 6. ACT Relief Package

package and invisible on the crisis map. The response process and a sample relief package are demonstrated in Figures 5 and 6. 5.3

Response Coordination

Relief organizations that respond independently can complicate relief efforts. If independent organizations duplicate a response, it will draw resources away from other needy areas that could use the duplicated supplies, delay response to other disaster areas, and/or result in additional transportation and security requirements. We base our approach to response coordination on the concept of “inter-agency [2]” to avoid response conflicts while maintaining the centralized control. To support response coordination, all available requests (i.e. requests that have not been filled) are displayed on the crisis map. Relief organizations look at available requests and selectively respond to the requests they can support. To avoid conflicts, relief organizations are not able to respond to requests another

Fig. 7. Request States Transition in ACT

Promoting Coordination for Disaster Relief

203

organization is fulfilling. Claimed requests that are not addressed for 24 hours become visible again on the crisis map. Each request is in one of four states: available, in process, in delivery, or delivered. After necessary preprocessing, submitted requests are put into the requests pool. Requests become available and visualized in the crisis map. If a request is selected by an organization, the state changes to in process. The request becomes available again if it is not addressed in 24 hours. The request state becomes in delivery if it is on the way to the final destination. Finally, the state becomes delivered if the request is fulfilled. A detailed state transition is summarized in Figure 7. 5.4

Statistics for Improved Communication

Relief organizations are usually not informed of other responders’ actions since they are only aware of the available requests. Statistics such as organizational contribution during relief operations are helpful in evaluating the relief progress, adjusting relief strategies and making further decisions. The following statistics give insights into the relief effort: – Current request delivery status. Information about current fulfillment status for each type of requests. – Spatio-temporal distribution of requests. Where and when information about requests and responses. – Distribution of contributed resource types of each organization. Each organization has its own specialization and strength in relief operations as well as different budgets on various resources. An example of distribution statistical reporting is shown in Figure 8. The left pie graph represents an example of detailed water requests, note that 10% of the requests are not fulfilled. The right bar graph represents the water delivery at a 10-day resolution in an earthquake for the Red Cross and United Nations.

Fig. 8. Statistics on water fulfillment and organizational contribution

204

6

H. Gao et al.

Conclusion and Future Work

Social media is being used as an efficient communications mechanism during disasters. In this work, we discussed the challenges with coordination for disaster relief and present an architecture aimed at improving disaster response beyond current practice. We believe this approach will enable organizations to cooperate with each other more efficiently. We are developing an open system, ACT, with the primary goal of providing relief organizations the means for better collaboration and coordination during a crisis. The system implementation is ongoing and additional work needs to be done to test our approach. We are investing approaches to enable collaboration and provide appropriate security to relief organizations and workers. We are designing reporting functions that are insightful for disaster relief. By both leveraging crowdsource information and providing the means for a groupsource response, organizations will be able to assist persons affected by a crisis more effectively in the future.

Acknowledgments This work is supported, in part, by the Office of Naval Research.

References 1. Dynes, R.R.: Organized Behavior in Disaster: Analysis and Conceptualization. US Dept. of Commerce, National Bureau of Standards (1969) 2. Goolsby, R.: Social media as crisis platform: The future of community maps/crisis maps. ACM Trans. Intell. Syst. Technol. 1(1), 1–11 (2010) 3. Howe, J.: The rise of crowdsourcing. Wired Magazine 14(6), 1–4 (2006) 4. Knabb, R.D., Rhome, J.R., Brown, D.P.: Tropical cyclone report: Hurricane katrina. National Hurricane Center 20 (August 23-30, 2005) 5. Liu, S.B., Palen, L.: Spatiotemporal mashups: A survey of current tools to inform next generation crisis support. In: Proceedings of the 6th International Conference on Information Systems for Crisis Response and Management, ISCRAM 2009 (2009) 6. Samaraweera, I., Corera, S.: Sahana victim registries: Effectively track disaster victims. In: Proceedings of the 4th International Conference on Information Systems for Crisis Response and Management, ISCRAM 2007 (2007)

Trust Maximization in Social Networks Justin Zhan and Xing Fang Department of Computer Science North Carolina A&T State University {Zzhan,xfang}@ncat.edu

Abstract. Trust is a human-related phenomenon in social networks. Trust research on social networks has gained much attention on its usefulness, and on modeling propagations. There is little focus on finding maximum trust in social networks which is particularly important when a social network is oriented by certain tasks. In this paper, we propose a trust maximization algorithm based on the task-oriented social networks. Keywords: Social Networks; Trust; Maximization.

1 Introduction With the advent of web 2.0 technologies, social networks are propagating rapidly. Instead of granting users passive browsing web content, it allows users to have personal accounts. The users may post their own words, pictures, and movies, etc. to the social network sties. Users on the social networks can also write comments, share personal preferences, and make friends with people they want. Examples of such networks are Facebook, MySpace, Twitter, and other blogs or forums. While more and more people are registering accounts on social networks and enjoying the new technologies [1], problems inherent with this not-well developed network environment finally came into reality. Hogben [2] pointed out that Trust (Trustworthy Users) will be one major security issue for the future social networks. Trust has been being a topic that intrigues intensive research work from numerous computational scientists. Owing to its human-related characteristics, trust is hard to be modeled or even uniformly defined for computational systems. Marsh introduced a trust computing model [3]. Unfortunately, this frequently-referenced work is highly theoretical and difficult to apply, particularly on social networks [4]. Sztompka [5] defined trust as a bet about the future contingent actions of others. Grandison and Sloman [6] addressed that trust is the firm belief in the competence of an entity to act dependably, securely, and reliably within a specified context. Mui et al. claimed that trust is a subjective expectation an agent has about another’s future behavior based on the history of their encounters [7]. Olmedilla et al. presented that trust of a party A to a party B for a service X is the measurable belief of A in that B behaves dependably for a specified period within a specified context (in relation to service X) [8]. As an important property, trust has also been involved in different types of networks. Online shopping networks such as Amazon and eBay allow buyers to evaluate sellers by posting comments and trust ratings after every purchase. Online J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, pp. 205–211, 2011. © Springer-Verlag Berlin Heidelberg 2011

206

J. Zhan and X. Fang

P2P lending network, like prosper.com [9], is based on mutual trust [10]. Trust is also served as a vital factor for selecting helpers in social authentication [11]. In addition to focusing on direct trust, trust inference mechanisms are introduced to infer an accurate trust value between two people who may not have direct connections (communications) on social networks [12]. Kamvar et al. proposed the Eigen Trust Algorithm, where the idea is to eliminate malicious peers on P2P networks through reputation management [13]. The algorithm is able to calculate a global trust value for each peer on the networks based on each peer’s historical behavior. The ones with relatively low values will be identified as malicious peers and will then be isolated. Similar approaches were adopted in [14, 15] to calculate trust values for the users on the semantic web and the trust-based recommendation system, respectively. The remainder of this paper is organized as follows: in section 2, we reviewed existing trust inference mechanisms. In section 3, we present the trust maximization algorithm based on task-oriented social networks.

2 Trust Inference Mechanisms Trust Inference Mechanism is designed to calculate an accurate trust value between two persons who may not have direct connections on social network. Liu et al. [16] categorize existing Trust Inference Mechanisms into three different types: Multiplication, Average, and Probabilistic mechanisms. 2.1 Multiplication Mechanism Trust transitivity and asymmetry are two major attributes for trust computation [4]. If person A highly trusts person B with value TAB and person B highly trusts person C with value TBC, A then is able to trust person C, to some extent, with value TAC [17, 18]. Thus trust transitivity is represented as: TAC = TAB * TBC. While trust asymmetry describes the truth that A trusts B (C) with value TAB (TAC) does not indicate B (C) trusts A with the same value. Titled as multiplication, the first type of trust inference mechanism is theoretically originated from trust transitivity as well as asymmetry. Ziegler and Lausen [19] propose Appleseed, a local group trust computation metric, to examine trust propagation via local social networks. Bhuiyan et al. [20] present a framework integrated with trust among community members and public reputation information for recommender systems. Zuo et al. [21] introduce a transitive trust evaluation framework to evaluate trust value for social network peers. 2.2 Average Mechanism Golbeck and Hendler [4] present an average trust inference mechanism. Trust value between two indirect connected nodes, source and sink, can be calculated via averaging the trust ratings to the sink that belong to the good nodes in between of the two. They believe in a graph with n nodes, probability of the bulk of the nodes will precisely rate the sink is following a binomial distribution,

a 1

a

, where n denotes the

number of the nodes and a denotes the accuracy that the nodes can correctly rate the sink. In this case, the magnitude of trust value actually depends on n and a.

Trust Maximization in Social Networks

207

2.3 Probabilistic Mechanism The third type of trust inference mechanism belongs to the probabilistic method. Kuter and Golbeck [22] proposed a Probabilistic Confidence Model for the trust value computing. The structure of this confidence model is based on a Bayesian Network in which the confidence values, which are used to estimate the trust values, are introduced and their bounds are calculated. Liu et al. [16] propose a trust inference mechanism based on trust-oriented social networks. In their trust inference mechanism, they introduce two variables r (0 r 1) and ρ (0 ρ 1). r is an intimate degree value of social relationships between participants in trust-oriented social networks. If rAB = 0, it indicates that A and B have no social relationship. rAB = 1 indicates A and B have the most intimate social relationship. ρ is a role impact value. It illustrates the impact of the participants’ recommendation roles (expert or beginner) in a specific domain in trust-oriented social networks. If ρA = 0, it indicates that A has no knowledge in the domain. If ρA = 1, it indicates A is a domain expert.

3 Task-Oriented Social Network and Trust Maximization Alg A task-oriented social network is able to accomplish certain tasks based on its formation. A general social network is defined as a connected graph G V, E , where V v , v , … , v is a set of n nodes, and E V V is a set of edges [23]. For each v V, it denotes an individual located on the social network. For each pair of u, v E, it denotes the connection (communication) between two nodes. A a ,a ,…,a is a universe of m skills. By denoting a v , we claim that the individual i has the skill j. And each individual is associated with a small set of skills, A. A task is defined as a subset of skills required to accomplish the where A v task, that is A. For a certain task, that can be performed by a group of individuals, V ′ , where V ′ V. Hence, this group of individuals forms a task-oriented V′, E′ . social network, G′ , where G′ G and G′ 3.1 Task-Oriented Social Network Generation (TOSNG) PROBLEM 1: /TOSNG/ Given a general social network , and a task , find ′ ′ , ′ that . a task-oriented social network ′ The TOSNG problem is essentially a Team Formation (TF) problem where a formed team aims to accomplish certain tasks or missions [23]. Several conditions that are able to affect the formation process are considered. The conditions may include individuals’ drive, temperament [24] and their personalities [25]. Moreover, communication-cost is introduced as a constraint of the TF problem [23] where Lappas et al. explored the computational complexity of solving the TF problem which has been shown NP-complete. 3.2 Maximum Trust Routes Identification (MTRI) PROBLEM 2: /MTRI/ Given a task-oriented social network collection of edges consisting of the maximum trust routes.

′

′

,

′

,

′

is a

208

J. Zhan and X. Fang

The aforementioned trust inference mechanisms provide fundamentals for solving the MTRI problem. However, one concern is their time complexity. We adopt the simplest multiplication method as our solution for the MTRI problem. Recall that trust routes are directed and trust values are vectors. In order to identify the maximum trust routes for a certain social network, we need to map the network to trust realm.

Fig. 1. Social Realm and Trust Realm

Figure 1 depicts a small social network consisting of three individuals Alice (A), Bob (B), and Carol (C). In social realm, A, B, and C are three nodes connected by undirected (normal) edges indicating that they have communications with each other. According to our previous research, trust value can be deduced based on the historical communications [12]. Hence, in a trust realm, the directed edges are weighted with certain trust values to demonstrate the trust relationships. For instance, |T | T , where T 0,1 , in which one means 100% trust and 0% means no trust at all.

Fig. 2. An Example of Trust Edges

In Figure 2, suppose that we want to identify the maximum trust route, based on the perspective of Alice (node A), and the identified route must contain Carol (node C). To tackle the problem, we first sort out all of the possible routes from A to C. Two routes, R and R , are identified. Second, trust value for each route is inferred by the multiplication method: 0.72 0.85 0.54

0.61

Trust Maximization in Social Networks

209

Given T 0.61 T 0.54 , R is then identified as the maximum trust route with removed directions showing in Figure 3. This example also reveals a very important property of trust routes. That is trust routes are not necessary the shortest routes.

Fig. 3. Maximum Trust Route in Social Realm

Trust-Maximization Algorithm Algorithm 1. The Trust-Maximization algorithm Input: , and ′ Output: Team ′ and ′ , 1: for do 2: | 3: argmin| 4: for do 5: for and do 6: MaxTrustRoute , ∑ TrustValue 7: Tr | 8: argmax|Tr ′ 9: ′ , ′

′

Algorithm 1 is the pseudocode of the Trust-Maximization algorithm. The algorithm is a variation of the RarestFirst algorithm in [23], for the purpose of team formation. In this sense, for each skill, a , required by a given task, , a support group S a is computed. The support group with the lowest cardinality is denoted by S a . MaxTrustRoute v , S a function returns the maximum trust route from v to S a , one support group. TrustValue function identifies the trust values of the maximum trust routes. v is one of the nodes belonging to S a that the routes starts from v totally have the maximum magnitued of trust. Suppose that all routes between any pair of nodes have been pre-identified. We also assume that each route’s trust value, on all directions, has been pre-computed. The running time complexity of the Trust-Maximization algorithm is n at the worst case. PROPOSITION: The output of the Trust-Maximization algorithm, of maximum trust routes, from the perspective of .

′

′

,

′

consists

PROOF: ′ is a task oriented social network that is able to accomplish the given task, . is a required skill from that is possessed by the least number of individuals comparing with any other required skills from . is a group of individuals

210

J. Zhan and X. Fang

who possess . It is intuitively appropriate to deploy trust routes from the perspective of because of its indispensability. is, at least, one individual from . are groups with individuals possessing the task-required skills . There may exist multiple routes, from to every . except MaxTrustRoute only returns the maximum trust route, from to every . We which is a collection of all the maximum trust routes from to then have . Recall that, for each route, its trust value is located in the domain of 0,1 . ∑ TrustValue is meant to add all of the trust values of the maximum trust routes from , the result will be a maximized trust value of . If is one individual, the output graph is the collection of ’s maximum trust routes to . If is multiple individuals, the output graph is also the collection of the maximum trust □ routes from the one who possess the highest maximized trust value.

References 1. Schonfeld, E.: Facebook Connect + Facebook Ads = A Social Ad Network, http://techcrunch.com/2009/03/03/ facebook-connect-facebook-ads-a-social-ad-network/ 2. Hogben, G.: Security Issues in the Future of Social Networking. In: Proceedings of W3C Workshop on the Future of Social Networking, Barcelona, Spain (January 2009) 3. Marsh, S.: Formalising Trust as a Computational Concept, PhD thesis, Department of Mathematics and Computer Science, University of Stirling (1994) 4. Golbeck, J., Hendler, J.: Inferring Trust Relationships in Web-based Social Networks. ACM Transactions on Internet Technology 6(4), 497–529 (2006) 5. Sztompka, P.: Trust: A Sociological Theory. Cambridge University Press, Cambridge (1999) 6. Grandison, T., Sloman, M.: A Survey of Trust in Internet Applications. Communications Surveys and Tutorials 4(4), 2–16 (2000) 7. Mui, L., Mohtashemi, M., Halberstadt, A.: A Computational Model of Trust and Reputation for E-business. In: Proceedings of the 35th International Conference on System Science, Big Island, Hawaii, pp. 280–287 (January 2002) 8. Olmedialla, D., Rana, O., Matthews, B., Nejdl, W.: Security and Trust Issues in Semantic Grids. In: Proceedings of the Dagsthul Seminar, Semantic Grid: The Convergence of Technologies, vol. 05271 9. Prosper, http://www.prosper.com/ 10. Shen, D., Krumme, C., Lippman, A.: Follow the Profit or the Herd? Exploring Social Effects in Peer-to-Peer Lending. In: Proceedings of The Second International Conference on Social Computing, Minneapolis, MN, USA, pp. 137–144 (August 2010) 11. Brainard, J., Juels, A., Rivest, R., Szydlo, M., Yung, M.: Fourth-Factor Authentication: Somebody You Know. In: Proceedings of the 13th ACM Conference on Computer and communications security, Virginia (October 2006) 12. Zhan, J., Fang, X.: A Computational Trust Framework for Social Computing. In: Proceedings of The Second International Conference on Social Computing, Minneapolis, MN, USA, pp. 264–269 (August 2010) 13. Kamvar, S., Schlosser, M., Garcia-Molina, H.: The Eigen Trust Algorithm for Reputation Management in P2P Networks. In: Proceedings of the 12th International World Wide Web Conference, Budapest, Hungary, pp. 640–651 (May 2003)

Trust Maximization in Social Networks

211

14. Richardson, M., Agrawal, R., Domingos, P.: Trust Management for the Semantic Web. In: Proceedings of the 2nd International Semantic Web Conference, Sanibel, Finland, pp. 351–368 (October 2003) 15. Walter, F., Battistion, S., Schweitzer, F.: A Model of A Trust-based Recommendation System. Auton Agent Multi-Agent System 16(1), 57–74 (2007) 16. Liu, G., Wang, Y., Orgun, M.: Trust Inference in Complex Trust-oriented Social Networks. In: Proceedings of the International Conference on Computational Science and Engineering, Vancouver, Canada, pp. 996–1001 (August 2009) 17. Jonsang, A., Gray, E., Kinaterder, M.: Simplification and Analysis of Transitive Trust Networks. Web Intelligence and Agent Systems 4(2), 139–161 (2006) 18. Golbeck, J.: Combining Provenance with Trust in Social Networks for Semantic Web Content Filtering. In: Proceedings of the International Provenance and Annotation Workshop, Chicago, Illinois, USA, pp. 101–108 (May 2006) 19. Ziegler, C., Lausen, G.: Propagation Models for Trust and Distrust in Social Networks. Information Systems Frontiers 7(4-5), 337–358 (2005) 20. Bhuiyan, T., Xu, Y., Josang, A.: Integrating Trust with Public Reputation in Locationbased Social Networks for Recommendation Making. In: Proceedings of the IEEE/WIC/ACM International Conference on Web Intelligence and Intelligent Agent Technology, Sydney, Australia, pp. 107–110 (December 2008) 21. Zuo, Y., Hu, W., Keefe, T.: Trust Computing for Social Networking. In: Proceedings of the 6th International Conference on Information Technology New Generations, Las Vegas, Nevada, USA, pp. 1534–1539 (April 2009) 22. Kuter, U., Golbeck, J.: SUNNY: A New Algorithm for Trust Inference in Social Networks Using Probabilistic Confidence Models. In: Proceedings of the 22nd AAAI Conference on Artificial Intelligence, British Columbia, Canada, pp. 1377–1382 (July 2007) 23. Lappas, T., Liu, K., Terzi, E.: Finding a Team of Experts in Social Networks. In: Proceedings of ACM International Conference on Knowledge Discovery and Data Mining (KDD 2009), Paris, France, pp. 467–475 (June-July 2009) 24. Fitzpatrick, E., Askin, R.: Forming Effective Worker Teams with Multi-Funcitional Skill Requirements. Computers and Industrial Engineering 48(3), 593–608 (2005) 25. Chen, S., Lin, L.: Modeling Team Member Characteristics for the Formation of a Multifunctional Team in Concurrent Engineering. IEEE Transactions on Engineering Management 51(2), 111–124 (2004)

Towards a Computational Analysis of Status and Leadership Styles on FDA Panels David A. Broniatowski and Christopher L. Magee Engineering Systems Division, Massachusetts Institute of Technology, E38-450, 77 Massachusetts Ave., Cambridge, MA 02141, USA {david,cmagee}@mit.edu

Abstract. Decisions by committees of technical experts are increasingly impacting society. These decision-makers are typically embedded within a web of social relations. Taken as a whole, these relations define an implicit social structure which can influence the decision outcome. Aspects of this structure are founded on interpersonal affinity between parties to the negotiation, on assigned roles, and on the recognition of status characteristics, such as relevant domain expertise. This paper build upon a methodology aimed at extracting an explicit representation of such social structures using meeting transcripts as a data source. Whereas earlier results demonstrated that the method presented here can identify groups of decision-makers with a contextual affinity (i.e., membership in a given medical specialty or voting clique), we now can extract meaningful status hierarchies, and can identify differing facilitation styles among committee chairs. Use of this method is demonstrated on the transcripts of U.S. Food and Drug Administration (FDA) advisory panel meeting transcripts; nevertheless, the approach presented here is extensible to other domains and requires only a meeting transcript as input. Keywords: Linguistic analysis, Bayesian inference, committee decision-making.

1 Introduction This paper presents a computational methodology designed to study decision-making by small groups of technical experts. Previously, we showed how a computational approach could be used to generate social networks from meeting transcripts [1]. The advantage of using a computational approach over more traditional social science content analysis methods include the repeatability and consistency of the approach, as well as time- and labor-saving advances in automation. Quinn et al. [2] provide a compelling justification for the adoption of a computational analysis by scholars of agenda-setting in political science. We feel that this argument may equally be extended to other domains of interest to social scientists and to social computing generally.

2 Data Source The empirical analysis mentioned above requires data in the form of committee meeting transcripts. The ideal data source must involve a negotiation (e.g., analysis or J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, pp. 212–218, 2011. © Springer-Verlag Berlin Heidelberg 2011

Towards a Computational Analysis of Status and Leadership Styles on FDA Panels

213

evaluation of a technological artifact); include participation of representatives from multiple groups (e.g., multiple experts from different fields or areas of specialization); contain a set of expressed preferences per meeting (such as a voting record); and multiple meetings must be available, so as to enable statistical significance. These requirements are met by the Food and Drug Administration’s medical device advisory panels – bodies that review the most uncertain medical devices prior to their exposure to the American market. A device’s approval and future market diffusion often rests upon the panel’s assessment of the device’s safety. These panels are aimed at producing a recommendation, informed by the expertise and knowledge of panel members, which can supplement the FDA’s “in-house” decision process. Multiple experts are consulted so that the group decision’s efficacy can take advantage of many different institutional perspectives. Panel members’ values and institutional contexts may differ, leading to different readings of the evidence, and therefore different recommendations. In health care, Gelijns et al. [3] note that strictly evidence-based decisions are often not possible due to different interpretations of data, uncertain patterns of technological change, and differing value systems among experts.

3 Methodology We use Bayesian topic models to determine whether actors within a committee meeting are using similar terminology to discuss the common problem to be solved. The choice of Bayesian topic models is driven by a desire to make the assumptions underlying this analysis minimal and transparent while maintaining an acceptable level of resolution of patterns inferred from the transcript data. We assume that each speaker possesses a signature in his or her word choice that can inform the selection of topics. We therefore use a variant of the Author-Topic (AT) Model [4], which creates probabilistic pressure to assign each author to a specific topic. Shared topics are therefore more likely to represent common jargon. Briefly, a social network is created by first generating a term-document matrix from a meeting transcript. Words are stemmed using Porter’s Snowball algorithm [5], and function-words are removed. The AT Model is then used to assign each word token to a topic. Following the procedure outlined in [1], each utterance by a committee voting member is assigned to both that individual and to a “false author”. Practically, this has the effect of removing procedural words from the analysis, and focusing instead on a speaker’s unique language. A collapsed Gibbs sampler, using a variant of the algorithm outlined in [6], is used to generate samples from the AT Model’s posterior distribution. The number of topics is chosen independently for each transcript as follows: 35 AT Models are fit to the transcript for T = 1 … 35 topics. An upper limit of 35 topics was chosen based upon empirical observation (see Broniatowski and Magee [1] for more details). For each model, 20 samples are generated from one randomly initialized Markov chain after an empirically-determined burn-in of 1000 iterations. We find the smallest value, t0, such that the 95th percentile of all samples for all larger values of T is greater than the 5th percentile of t0. We set the value of T = t0 + 1 so as to ensure that the model chosen is beyond the knee in the curve.

214

D.A. Broniatowski and C.L. Magee

Once the number of topics has been chosen, a T-topic AT Model is again fit to the transcript. Following the procedure used by Griffiths and Steyvers [6], ten samples are taken from 20 randomly initialized Markov chains, such that there are 200 samples in total. These form the basis for all subsequent analysis. A network is generated for each sample by examining the joint probability distribution for each author-pair in that sample: T

P(X 1 ∩ X 2 ) =

∑ P (Z = z

i

| X 1 ) * P (Z = z i | X 2 )

∑ P (Z = z

i

| X 1 ) * P (Z = z j | X 2 )

i T ,T i, j

If the joint probability distribution exceeds 1/T, then we say that this author-pair is linked in this sample. Using the Bonferroni criterion for a binomial distribution with family-wise error rate of 5%, and assuming ~15 voting committee members, two authors are connected by an edge in the transcript-specific social-network, Δ, if they are linked more than 125 times out of 200 samples. Full details are given by Broniatowski and Magee [1], where we showed that this approach can be used to identify groups of speakers who share similar identities (e.g., members of the same medical specialty or voting clique). We extend this method such that directed edges are overlaid on top of the social network. Directionality was determined by examining the order in which each speaker discussed a particular topic. For example, if two speakers, X and Y, share a topic and X consistently speaks after Y does, then we say that Y leads X on this topic. Conversely, X lags Y on this topic. These directions are then averaged over all topics in proportion to the amount that each speaker discusses each topic. In particular, consider a sample, s, from the posterior distribution of the AT model. Within this sample, choose a pair of speakers, x1 and x2, and a topic z. Given that utterances are temporally ordered, this defines two separate time-series. These time series can be used to generate the topic-specific cross correlation for speakers x1 and x2, in topic z:

( f i ,sz ∗ f js, z )[δ ] =

∞

∑f

d = −∞

s* i,z

[ d ] f js, z [δ + d ]

fi,ts

where (d) is the number of words spoken by author i and assigned to topic z in document d, in sample s. For each sample, s, from the AT Model’s posterior distribution, we examine the cross-correlation function for each author pair, {xi, xj}, in topic z. Let there be k peaks in the cross-correlation function, such that

mk = arg max ( f i t ∗ f jt )[δ ] δ

For each peak, if mk > 0, we say that author i lags author j in topic z, at point mk (i.e.,

l is, j , z , m k = 1 ). Similarly, we say that author i leads author j in topic z at point mk (i.e.,

l is, j , z , mk = −1 ) if mk < 0. Otherwise, l is, j , z , mk = 0 . For each sample, s, we define the polarity of authors i and j in topic z to be the median of the lsi,j,zt.

p is, j ,t = median(l is, j , z )

Towards a Computational Analysis of Status and Leadership Styles on FDA Panels

215

If most of the peaks in the cross-correlation function are greater than zero, then the polarity = 1; if most of the peaks are less than zero, then the polarity = -1; otherwise, the polarity = 0. We are particularly interested in the topic polarities for author-pairs who are linked in the graph methodology outlined above – i.e., where Δi,j = Δj,i =1. Using the polarity values defined above, we are interested in determining directionality in Δ. For each sample, s, we define the direction of edge ei,j in sample s as: T

d s (e i , j ) = ∑ ( p is, j ,t * P s ( Z = z i | x i ) * P s ( Z = z i | x j )) t =1

This expression weights each topic polarity by its importance in the joint probability distribution between xi and xj, and is constrained to be between -1 and 1 by definition. The set of 200 ds(ei,j) defines a distribution such that the net edge direction, d(ei,j) is determined by partition of the unit interval into three segments of equal probability. In particular, this is done by examining the proportion of the ds(ei,j) that are greater than 0. If more than 66% of the ds(ei,j) > 0 then d(ei,j) = 1 (the arrow points from j to i). If less than 33% of ds(ei,j) > 0 then d(ei,j) = -1 (the arrow points from i to j). Otherwise, d(ei,j) = 0 (the arrow is bidirectional). The result is a directed graph, an example of which is shown in Figure 1. It is interesting to note that these graphs may change their topology on the basis of which nodes are included. In particular, inclusion of the committee chair may or may not "flatten" the graph by connecting the sink nodes (i.e., nodes on the bottom of the graph) to higher nodes. In other meetings, the chair may act to connect otherwise disparate groups of voters or serve to pass information from one set of nodes to another.

Fig. 1. Directed Graph representation of meeting held on June 23, 2005. Hierarchy metric = 0.35 [10]. Node shape represents medical specialty and note color represents vote (white is approval; grey is non-approval; black is abstention).

216

D.A. Broniatowski and C.L. Magee

Fig. 2. Directed Graph representation of meeting held on June 23, 2005, with the committee chair included. Hierarchy metric = 0.78. Node shape represents medical specialty and note color represents vote (white is approval; grey is non-approval; black is abstention).

q Voting Member 4

Voting Member 2

Voting Member 6

Voting Member 5 Voting Member 7 Voting Member 8

Voting Member 1

Voting Member 3

Voting Member 4

Voting Member 5 Voting Member 7 Voting Member 8

Voting Member 1

Committee Chair

Voting Member 2

Voting Member 3 Voting Member 6

Fig. 3. Directed Graph representation of meeting held on October 27, 1998 without (left) and with (right) the committee chair. Hierarchy metric = 0 for both cases. Node shape represents medical specialty and note color represents vote (white is approval; grey is non-approval; black is abstention).

Towards a Computational Analysis of Status and Leadership Styles on FDA Panels

217

We may quantify the impact that the committee chair has upon the meeting by determining, for each graph, which proportion of edges is part of a cycle. This is a metric of the hierarchy in the graph [10]. The difference in this metric between graphs with and without the chair therefore quantifies impact of the chair on the meeting. We display this metric for the graph without the chair (e.g., Figure 2) and with the chair (e.g., Figure 3). The difference in this metric between graphs with and without the chair therefore quantifies the impact of the chair on the meeting. For the meeting held on June 23, 2005, this value is 0.78-0.35 = 0.43. This suggests that the chair is significantly changing the topology of the meeting structure – in particular, it seems that the chair is connecting members at the “bottom” of the graph to those at the “top”. Other meetings display different behavior by the chair. Consider the meeting held on October 27, 1998 (in Figure 3). In this meeting, the committee chair served to connect the two disparate clusters on the panel. Nevertheless, the chair is not creating any new cycles on the graph. This is reflected in the fact that the hierarchy metric for both of these meetings is equal to 0. 5

7.325

x 10

36

37

35 33

32

Date (days since January 1, 1900)

7.32

30 27 7.315

24 25 23

7.31 16

20

22

17 March 4, 2002

7.305

7.3

7.295 -0.05

8 7 1 0

0.05

0.1

0.15 0.2 0.25 Chair Impact

0.3

0.35

0.4

0.45

Fig. 4. Impact of chair vs. meeting date for each of the 17 meetings in which there was a voting minority. Each meeting is labeled by its corresponding ID from the sample of all 37. Meetings with a chair impact greater than zero are significantly more likely to have occurred later than are meetings with a chair impact of zero or less (p=0.005, using a non-parametric Mann-Whitney U test).

The chair’s role in “flattening” the structure of a given meeting’s graph could suggest a particular facilitation strategy, wherein the committee chair tries to elicit the opinions of voting members who speak later, and might otherwise be less influential. When the chair does not act to significantly change the graph structure, the chair may be taking on the role of a synthesizer – gathering the opinions of the other voters to answer FDA questions, but not causing any of the other members to repeat what has

218

D.A. Broniatowski and C.L. Magee

already been said. This suggests that there may be at least two different strategies that committee chairs could use during a meeting. We note that the impact of the chair seems to be increase markedly around March 2002 (see Figure 4). Furthermore, meetings with a chair impact greater than zero are significantly more likely to have occurred later than are meetings with a chair impact of zero or less (p=0.005, using a non-parametric Mann-Whitney U test).

4 Conclusions The marked change in strategy across multiple chairs has many potential explanations, e.g., a change in FDA policy regarding committee operations, e.g., a change in conflict of interest reporting procedures, but there is no obvious connection between these two events. Alternatively, the types of devices that came to the Circulatory Systems Devices Panel seem to have changed around this time; perhaps there was an increase in the difficulty of the devices that the panel reviewed (e.g., concurrent with the entrance on the market of drug-eluting stents, Left Ventricular Assist Devices and other potentially risky devices). Finally, we might hypothesize some combination of these ideas – that a change in FDA policy might have some way impacted upon chair style, and that this change in policy might have been driven by changing market conditions. Testing these hypotheses requires looking across multiple panels and committees, which we leave to future work. Nevertheless, if these findings were cross-validated with other metrics of committee chair efficacy, the method presented here suggests the possibility for a fruitful analysis of leadership dynamics. Acknowledgements. The author would like to acknowledge the MIT-Portugal Program for its generous support of this research.

References 1. Broniatowski, D., Magee, C.L.: Analysis of Social Dynamics on FDA Panels using Social Networks Extracted from Meeting Transcripts. In: 2nd IEEE Conference on Social Computing, 2nd Symposium on Social Intelligence and Networking (2010) 2. Quinn, K.M., et al.: How to Analyze Political Attention with Minimal Assumptions and Costs. American Journal of Political Science 54, 209–228 (2010) 3. Gelijns, A.C., et al.: Evidence, Politics, And Technological Change. Health Affairs 24, 29– 40 (2005) 4. Rosen-Zvi, M., et al.: The author-topic model for authors and documents. In: Proceedings of the 20th Conference on Uncertainty in Artificial Intelligence, pp. 487–494 (2004) 5. Porter, M.F.: Snowball: A language for stemming algorithms (October 2001) 6. Griffiths, T.L., Steyvers, M.: Finding scientific topics. Proceedings of the National Academy of Sciences of the United States of America 101, 5228–5235 (2004) 7. Luo, J., et al.: Measuring and Understanding Hierarchy as an Architectural Element in Industry Sectors. SSRN eLibrary (2009), http://papers.ssrn.com/sol3/papers.cfm?abstract_id=1421439

Tracking Communities in Dynamic Social Networks Kevin S. Xu1 , Mark Kliger2 , and Alfred O. Hero III1 1

2

EECS Department, University of Michigan, 1301 Beal Avenue Ann Arbor, MI 48109-2122 USA {xukevin,hero}@umich.edu Medasense Biometrics Ltd., P.O. Box 633, Ofakim, 87516 Israel [email protected]

Abstract. The study of communities in social networks has attracted considerable interest from many disciplines. Most studies have focused on static networks, and in doing so, have neglected the temporal dynamics of the networks and communities. This paper considers the problem of tracking communities over time in dynamic social networks. We propose a method for community tracking using an adaptive evolutionary clustering framework. We apply the method to reveal the temporal evolution of communities in two real data sets. In addition, we obtain a statistic that can be used for identifying change points in the network. Keywords: dynamic, social network, community, tracking, clustering.

1

Introduction

Traditionally, social network data have been collected through means such as surveys or human observation. Such data provide a view of a social network as a static entity over time. However, most social networks are dynamic structures that evolve over time. There has been recent interest in analyzing the temporal dynamics of social networks, enabled by the collection of dynamic social network data by electronic means such as cell phones, email, blogs, etc. [2, 6, 10]. A fundamental problem in the analysis of social networks is the detection of communities [7]. A community is often defined as a group of network members with stronger ties to members within the group than to members outside of the group. Previous studies on the community structure of social networks have typically focused on static networks. In doing so, the temporal dynamics of the networks and communities have been neglected. The natural extension of community detection to dynamic networks is community tracking, which makes it possible to observe how communities grow, shrink, merge, or split with time. In this paper, we propose a method for tracking communities in dynamic social networks. The proposed method makes use of an evolutionary clustering framework that detects communities at each time step using an adaptively weighted combination of current and historical data. The result is a set of communities at each time step, which are then matched with communities at other time steps so J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, pp. 219–226, 2011. c Springer-Verlag Berlin Heidelberg 2011 

220

K.S. Xu, M. Kliger, and A.O. Hero III

that communities can be tracked over time. We apply the proposed method to reveal the temporal evolution of communities in two real data sets. The detected communities are more accurate and easier to interpret than those detected by traditional approaches. We also obtain a statistic that appears to be a good identifier of change points in the network.

2

Methodology

The objective of this study is to track the evolution of communities over time in dynamic social networks. We represent a social network by an undirected weighted graph, where the nodes of the graph represent the members of the network, and the edge weights represent the strengths of social ties between members. The edge weights could be obtained by observations of direct interaction between nodes, such as physical proximity, or inferred by similarities between behavior patterns of nodes. We represent a dynamic social network by a sequence of time snapshots, where the snapshot at time step t is represented t by W t = wij , the matrix of edge weights at time t. W t is commonly referred to as the adjacency matrix of the network snapshot. The problem of detecting communities in static networks has been studied by researchers from a wide range of disciplines. Many community detection methods originated from methods of graph partitioning and data clustering. Popular community detection methods include modularity maximization [7] and spectral clustering [12, 14]. In this paper, we address the extension of community detection to dynamic networks, which we call community tracking. We propose to perform community tracking using an adaptive evolutionary clustering framework, which we now introduce. 2.1

Adaptive Evolutionary Clustering

Evolutionary clustering is an emerging research area dealing with clustering dynamic data. First we note that is possible to cluster dynamic data simply by performing ordinary clustering at each time step using the most recent data. However this approach is extremely sensitive to noise and produces clustering results that are inconsistent with results from adjacent time steps. Evolutionary clustering combines data from multiple time steps to compute the clustering result at a single time step, which allows clustering results to vary smoothly over time. Xu et al. [13] recently proposed an evolutionary clustering framework that adaptively estimates the optimal weighted combination of current and past data to minimize a mean-squared error (MSE) criterion. We describe the framework in the following. Define a smoothed adjacency matrix at time t by   ¯ t = αt W ¯ t−1 + 1 − αt W t W (1) ¯ 0 = W 0 . αt can be interpreted as a forgetting factor that for t ≥ 1 and by W controls the amount of weight given to past data. We treat each network snapshot W t as a realization from a nonstationary random process and define the

Tracking Communities in Dynamic Social Networks

221

 t expected adjacency matrix Ψ t = ψij = E [W t ]. If we had access to the expected t adjacency matrix Ψ , we would expect to see improved clustering results by clustering on Ψ t rather than the noisy realization W t . However, Ψ t is unknown in real applications so the goal is to estimate it as accurately as possible. If we take the estimate to be the convex combination defined in (1), it was shown in [13] t that theoptimal choice  of α that minimizes the MSE in terms of the Frobenius t t 2 ¯ norm E W − Ψ F is given by n  n 

 t ∗ α =

n  n    i=1 j=1

 t var wij

i=1 j=1 t−1 w ¯ij

−

 t 2 ψij

+ var



t wij

, 

(2)

where n denotes in the network. In a real application,  t  the number of nodes ∗ t ψij and var wij are unknown so (αt ) cannot be computed. However, it can be approximated by replacing the unknown means and variances with sample means and variances. The communities at time t can then be extracted by performing ¯ t. ordinary community detection on the smoothed adjacency matrix W Any algorithm for ordinary community detection can be used with the adaptive evolutionary clustering framework. In this paper, we use Yu and Shi’s normalized cut spectral clustering algorithm [14]. It finds a near global-optimal separation of the nodes into k communities, where k is specified by the user. The algorithm involves computing the eigenvectors corresponding to the k largest ¯ t , then discretizing the eigenvectors so eigenvalues of a normalized version of W that each node is assigned to a single community. We refer readers to [13] for additional details on the adaptive evolutionary spectral clustering algorithm. 2.2

Tracking Communities over Time

There are several additional issues that also need to be addressed in order to track communities over time. The communities detected at adjacent time steps need to be matched so that we can observe how any particular community evolves over time. This can be achieved by finding an optimal permutation of the communities at time t to maximize agreement with those at time t − 1. If the number of communities at time t is small, it is possible to exhaustively search through all such permutations. This is, however, impractical for many applications. We employ the following heuristic: match the two communities at time t and t − 1 with the largest number of nodes in agreement, remove these communities from consideration, match the two communities with the second largest number of nodes in agreement, remove them from consideration, and so on until all communities have been exhausted. Another issue is the selection of the number of communities k at each time. Since the evolutionary clustering framework involves simply taking convex combinations of adjacency matrices, any heuristic for choosing the number of communities in ordinary spectral clustering can also be used in evolutionary spectral

222

K.S. Xu, M. Kliger, and A.O. Hero III

¯ t instead of W t . In this paper we use the eigenclustering by applying it to W gap heuristic [12] of selecting the number of communities k such that the gap between the kth and (k + 1)th largest eigenvalues of the normalized adjacency matrix is large. Finally, there is the issue of nodes entering or leaving the network over time. We deal with these nodes in the following manner. Nodes that leave the network ¯ t−1 in (1). Nodes that between times t − 1 and t can simply be removed from W ¯ t−1 . enter the network at time t have no corresponding rows and columns in W Hence, these new nodes can be naturally handled by adding rows and columns ¯ t after performing the smoothing operation in (1). In this way, the new to W nodes have no influence on the update of the forgetting factor αt yet contribute ¯ t. to the community structure through W

3 3.1

Experiments Reality Mining

Data Description. The MIT Reality Mining data set [2] was collected as part of an experiment on inferring social networks by monitoring cell phone usage rather than by traditional means such as surveys. The data was collected by recording cell phone activity of 94 students and staff at MIT for over a year. Each phone recorded the Media Access Control (MAC) addresses of nearby Bluetooth devices at five-minute intervals. Using this device proximity data, we construct a sequence of adjacency matrices where the edge weight between two participants corresponds to the number of intervals where they were in close physical proximity within a time step. We divide the data into time steps of one week, resulting in 46 time steps between August 2004 and June 2005. In this data set, we have partial ground truth to compare against. From the MIT academic calendar [5], we know the dates of important events such as the beginning and end of school terms. In addition, we know that 26 of the participants were incoming students at the university’s business school, while the rest were colleagues working in the same building. Thus we would expect the detected communities to match the participant affiliations, at least during the school terms when students are taking classes. Observations. We make several interesting observations about the community structure of this data set and its evolution over time. The importance of temporal smoothing for tracking communities can be seen in Fig. 1. On the left is the heat map of community membership over time when the proposed method is used. On the right is the same heat map when ordinary community detection at each time is used, which is equivalent to setting αt = 0 in (1). Notice that two clear communities appear in the heat map to the left, where the proposed method is used. The participants above the black line correspond to the colleagues working in the same building, while those below the black line correspond to the incoming business school students. On the heat map to the right, corresponding to ordinary

Tracking Communities in Dynamic Social Networks

20 Participant

Participant

20

223

40 60

40 60 80

80 10

20 30 Time step

10

40

20 30 Time step

40

Fig. 1. Heat maps of community structure over time for the proposed method (left) and ordinary community detection (right) in the Reality Mining experiment 1 Fall term begins

Fall term ends

Spring break begins

Estimated α

t

0.8 0.6 0.4 0.2 0

Winter term begins

Thanksgiving 0

5

10

15

20

25

Winter term ends 30

35

40

45

Time step

Fig. 2. Estimated forgetting factor αt by time step in the Reality Mining experiment

community detection, the community memberships fluctuate highly over time. Thus we can see that tracking communities by the proposed method results in more stable and accurately identified communities. The estimated forgetting factor αt at each time step is plotted in Fig. 2. Six important dates are labeled on the plot. Notice that the estimated forgetting factor drops several times, suggesting that the structure of the proximity network changes, around these dates. This is a reasonable result because the proximity network should be different when students are not in school compared to when they are in school. Thus αt also appears to be a good identifier of change points in the network. 3.2

Project Honey Pot

Data Description. Project Honey Pot [8] is an ongoing project targeted at identifying spammers. It consists of a distributed network of decoy web pages with trap email addresses, which are collected by automated email address harvesters. Both the decoy web pages and the email addresses are monitored, providing us with information about the harvester and email server used for each spam email received at a trap address. A previous study on the Project Honey Pot data [9] found that harvesting is typically done in a centralized manner. Thus harvesters are likely to be associated with spammers, and in this study we assume that the harvesters monitored by Project Honey Pot are indeed representative of spammers. This allows us to associate each spam email with a spammer so that we can track communities of spammers.

224

K.S. Xu, M. Kliger, and A.O. Hero III

2

1000 1500

20

4

Spammer

Spammer

Spammer

500

6 8 10

2000 4

8 6 Time step

10

12

60 80 100

12

2

40

2

4

8 6 Time step

10

12

2

4

8 6 Time step

10

12

Fig. 3. Temporal evolution of the giant community (left), a persistent community (middle), and a “staircase” community (right) in the Project Honey Pot experiment

Unlike in the previous experiment, we cannot observe direct interactions between spammers. The interactions must be inferred through indirect observations. We take the edge weight between two spammers i and j to be the total number of emails sent by i and j through shared email servers, normalized by the product of the number of email addresses collected by i and by j. Since servers act as resources for spammers to distribute emails, the edge weight is a measure of the amount of resources shared between two spammers. We divide the data set into time steps of one month and consider the period from January 2006 to December 2006. The number of trap email addresses monitored by Project Honey Pot grows over time, so there is a large influx of new spammers being monitored at each time step. Some spammers also leave the network over time. Observations. In this data set, we do not have ground truth for validation so the experiment is of an exploratory nature. At each time step, there are over 100 active communities, so rather than attempting to visualize all of the communities, as in Fig. 1, we instead try to visualize the evolution of individual communities over time. We discover several interesting evolution patterns, shown in Fig. 3. On the left, there is a giant community that continually grows over time as more and more spammers enter the network. The appearance of this giant community is quite common in large networks, where a core-periphery structure is typically observed [4]. In the middle, we illustrate a community that is persistent over time. Notice that no spammers change community until time step 12, when they all get absorbed into the giant community. Perhaps the most interesting type of community we observe is pictured on the right. We call this a “staircase” community due to the shape of the heat map. Notice that at each time step, many new spammers join the community while some of the existing spammers become inactive or leave the community. This suggests that either the members of the community are continually changing or that members assume multiple identities and are using different identities at different times. Since spamming is an illegal activity in many countries, the latter explanation is perhaps more likely because it makes spammers more difficult to track due to the multiple identities. Using the proposed method, it appears that we can indeed track these types of spammers despite their attempts to hide their identities.

Tracking Communities in Dynamic Social Networks

4

225

Related Work

There have been several other recent works on the problem of tracking communities in dynamic social networks. [11] proposed to identify communities by graph coloring; however, their framework assumes that the observed network at each time step is a disjoint union of cliques, whereas we target the more general case where the observed network can be an arbitrary graph. [3] proposed a method for tracking the evolution of communities that applies to the general case of arbitrary graphs. The method involves first performing ordinary community detection on time snapshots of the network by maximizing modularity. A graph of communities detected at each time step is then created, and meta-communities of communities are detected in this graph to match communities over time. The main drawback of this approach is that no temporal smoothing is incorporated, so the detected communities are likely to be unstable. Other algorithms for evolutionary clustering have also been proposed. Relevant algorithms for community tracking include [6] and [1], which extend modularity maximization and spectral clustering, respectively, to dynamic data. [10] proposed an evolutionary spectral clustering algorithm for dynamic multi-mode networks, which have different classes of nodes and interactions. Such an algorithm is particularly interesting for data where both direct and indirect interations can be observed. However, one shortcoming in these algorithms is that they require the user to determine to choose the values for parameters that control how smoothly the communities evolve over time. There are generally no guidelines on how these parameters can be chosen in an optimal manner.

5

Conclusion

In this paper, we introduced a method for tracking communities in dynamic social networks by adaptive evolutionary clustering. The method incorporated temporal smoothing to stabilize the variation of communities over time. We applied the method to two real data sets and found good agreement between our results and ground truth, when it was available. We also obtained a statistic that can be used for identifying change points. Finally, we were able to track communities where the members were continually changing or perhaps assuming multiple identities, which suggests that the proposed method may be a valuable tool for tracking communities in networks of illegal activity. The experiments highlighted several challenges that temporal tracking of communities presents in addition to the challenges present in static community detection. One major challenge is in the validation of communities, both with and without ground truth information. Another major challenge is the selection of the number of communities at each time step. A poor choice for the number of communities may create the appearance of communities merging or splitting when there is no actual change occurring. This remains an open problem even in the case of static networks. The availability of multiple network snapshots may actually simplify this problem since one would expect that the number of communities, much like the community memberships, should evolve smoothly

226

K.S. Xu, M. Kliger, and A.O. Hero III

over time. Hence, the development of methods for selecting the number of communities in dynamic networks is an interesting area of future research. Acknowledgments. This work was partially supported by the National Science Foundation grant number CCF 0830490. Kevin Xu was supported in part by an award from the Natural Sciences and Engineering Research Council of Canada. The authors would like to thank Unspam Technologies Inc. for providing us with the Project Honey Pot data.

References 1. Chi, Y., Song, X., Zhou, D., Hino, K., Tseng, B.L.: Evolutionary spectral clustering by incorporating temporal smoothness. In: Proc. 13th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (2007) 2. Eagle, N., Pentland, A., Lazer, D.: Inferring friendship network structure by using mobile phone data. Proceedings of the National Academy of Sciences 106(36), 15274–15278 (2009) 3. Falkowski, T., Bartelheimer, J., Spiliopoulou, M.: Mining and visualizing the evolution of subgroups in social networks. In: Proc. IEEE/WIC/ACM International Conference on Web Intelligence (2006) 4. Leskovec, J., Lang, K.J., Dasgupta, A., Mahoney, M.W.: Statistical properties of community structure in large social and information networks. In: Proc. 17th International Conference on the World Wide Web (2008) 5. MIT academic calendar 2004-2005, http://web.mit.edu/registrar/www/calendar0405.html 6. Mucha, P.J., Richardson, T., Macon, K., Porter, M.A., Onnela, J.P.: Community structure in time-dependent, multiscale, and multiplex networks. Science 328(5980), 876–878 (2010) 7. Newman, M.E.J.: Modularity and community structure in networks. Proceedings of the National Academy of Sciences 103(23), 8577–8582 (2006) 8. Project Honey Pot, http://www.projecthoneypot.org 9. Prince, M., Dahl, B., Holloway, L., Keller, A., Langheinrich, E.: Understanding how spammers steal your e-mail address: An analysis of the first six months of data from Project Honey Pot. In: Proc. 2nd Conference on Email and Anti-Spam (2005) 10. Tang, L., Liu, H., Zhang, J., Nazeri, Z.: Community evolution in dynamic multimode networks. In: Proc. 14th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (2008) 11. Tantipathananandh, C., Berger-Wolf, T., Kempe, D.: A framework for community identification in dynamic social networks. In: Proc. 13th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (2007) 12. von Luxburg, U.: A tutorial on spectral clustering. Statistics and Computing 17(4), 395–416 (2007) 13. Xu, K.S., Kliger, M., Hero III, A.O.: Evolutionary spectral clustering with adaptive forgetting factor. In: Proc. IEEE International Conference on Acoustics, Speech, and Signal Processing (2010) 14. Yu, S.X., Shi, J.: Multiclass spectral clustering. In: Proc. 9th IEEE International Conference on Computer Vision (2003)

Bayesian Networks for Social Modeling Paul Whitney, Amanda White, Stephen Walsh, Angela Dalton, and Alan Brothers Pacific Northwest National Laboratory, Richland, WA USA 99354 {paul.whitney,amanda.white,stephen.walsh, angela.dalton,alan.brothers}@pnl.gov

Abstract. This paper describes a body of work developed over the past five years. The work addresses the use of Bayesian network (BN) models for representing and predicting social/organizational behaviors. The topics covered include model construction, validation, and use. These topics show the bulk of the lifetime of such model, beginning with construction, moving to validation and other aspects of model "critiquing", and finally demonstrating how the modeling approach might be used to inform policy analysis. The primary benefits of using a welldeveloped computational, mathematical, and statistical modeling structure, such as BN, are 1) there are significant computational, theoretical and capability bases on which to build 2) the ability to empirically critique the model, and potentially evaluate competing models for a social/behavioral phenomenon. Keywords: Social modeling, calibration, validation, diagnostics, expert elicitation, evidence assessment.

1 Introduction Computational modeling of human behavior (CMHB) is a rapidly growing enterprise. Drivers for this enterprise are both practical (prediction, understanding, and policy) and theoretical. We describe a collection of work in CMHB that is based on using Bayesian networks (BN) models [1] as the underlying computational engine. Computational approaches for CMHB include systems dynamics (SD), agentbased modeling (ABM), discrete event simulations (DES), and statistical methods. These modeling methods are used in dynamic settings. BN models are arguably best suited for semi-static phenomena or assessing over some time interval. Accordingly, we view BN and the dynamic approaches as complimentary. Regression analyses and related statistical methods are frequently used approaches for developing models. Relative to statistical approaches, the primary benefit of using BN-based modeling approaches is that models can be constructed in advance of data collection, and for phenomena for which data are or will not be available. The remainder of the paper reflects the life cycle of a BN model. The next section describes the construction of a BN model. Then we present technologies for calibrating the parameters and validating the structure of a BN model given empirical data. Finally, we illustrate how evidence can be attached to a BN through a notional example. J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, pp. 227–235, 2011. © Springer-Verlag Berlin Heidelberg 2011

228

P. Whitney et al.

2 Constructing Bayesian Network Models This section focuses on two fundamental aspects of model construction. The first is the methodology used to construct a model with information obtained from technical literature or direct discussions with domain experts. The second aspect is a methodology used to infer the model parameters based on domain experts’ feedback elicited from pair-wise scenario comparisons. 2.1 Model Construction We have executed the following approach to arriving at a BN model structure for a human or organization’s behavior. The approach takes as raw input materials from a social science domain. We execute the following steps: • • •

Key concepts influencing the phenomena of interest are identified Relationships among these concepts, as suggested by the literature, are identified A BN is constructed, based on the concepts identified, that is consistent with the relationships among the concepts.

While the three-step outline is simple, the execution can be challenging to various degrees, depending on the source literature and complexity of the concept relationships within that literature. A source literature example that leads rapidly to BN models is political radicalization, as described in [2]. This paper provides a list of mechanisms that can lead to radicalization at various scales (individual, organization, and "mass"). Each of these in turn has influencing factors. One of the mechanisms involves the relationships – both connectedness and radicalization levels – between an individual and a group.

Fig. 1. Bayesian network model fragment representing one aspect of radicalization

Some of the concepts and their relationships within this source document include: • • •

If the individual is a member of a group, then that individual’s radicalization levels will be similar to the group’s If the individual is a recent member of the group, then their radicalization characteristics will be somewhere between the general population and that group If the individual is not a member of the group, their individual radicalization characteristics will be presumed to be similar to the larger population

A model fragment consistent with these relationships is shown in the Figure 1.

Bayesian Networks for Social Modeling

229

Both model structure and parameter values are needed for a BN. In this case, the parameters were set so that the above three relationships would hold. However, there are multiple BN structures and parameter values that could satisfy the above three relationships. The next subsection shows a method for mapping experts’ assessments of scenarios to parameter values. The subsequent section provides empirical methods for examining both the structure and model parameter values. 2.2 Experts’ Assessments Calibration Experts’ assessments play two prominent roles in BN model development: specifying the network structure, and providing parameters values for which pertinent data are lacking. The latter task presents considerable challenges in Bayesian network (BN) modeling due to heuristics and biases in decision making under uncertainty [3, 4]. The available data to calibrate a BN often provide limited coverage, or could be of poor quality or completely missing. Domain experts are sometimes the only viable information source. Developing a valid and efficient method to better capture domain experts’ insight is crucial for BN modeling. Providing assessments directly in probabilistic terms can be especially difficult for domain experts who are unfamiliar with probabilities. Elicitation approaches that demand less statistical esotericism are potentially useful for BN calibration. A popular method in marketing research, conjoint analysis [5] offers a novel applicability to elicitations for BN models. We developed a web-based interactive conjoint analysis expert elicitation user interface that offers pair-wise scenario comparisons for experts’ feedback as shown in Figure 2. The scenarios are combinations of nodes and node states from the underlying BN model. The expert indicates his/her assessment by moving the anchor along the scale, then submits the assessments, and immediately engages in another comparison. This process is run iteratively until sufficient information is collected for model calibration. Our initial test with conjoint analysisbased probability elicitation suggests that a fairly large amount of assessments can be obtained within a rather short time frame and that the elicitation approach has great potential for model calibration [6].

Fig. 2. Bayesian network model fragment representing one aspect of radicalization

3 Empirical Model Validation In the first sub-section, we describe how empirical observations are used in model building and validation. We define model calibration as estimating the model

230

P. Whitney et al.

parameters from empirical observations. In the second sub-section, we present a graphical diagnostic for the structural fit of the model with data. The major benefits of such a calibration include: 1. 2. 3. 4. 5.

A calibrated model that can, with some measurable level of precision, forecast observations within the same data space The ability to measure the quality of the model fit to the observations The ability to inspect the influence of more extreme observations on the model fit The ability to compare the relative fits of different hypothesized model structures The ability to measure and inspect the degree of consistency of the model structure.

The above capabilities are integral to testing domain theory against the state of nature and to arriving at a "good," trustable model that may be utilized in making prediction or other inference. 3.1 Empirical Model Calibration The Bayesian approach to calibration requires the specification of a prior distribution on the model parameters, Pr(θ), and may be combined with the data through the likelihood L(θ |X) to arrive at an updated posterior distribution Pr(θ |X) [7]. This is achieved through the general Bayesian estimation equation:

Pr(θ | X) ∝ L (θ | X) × Pr(θ ) . Point estimates and uncertainties are readily derived from the posterior. The arguments to adopt a Bayesian approach to parameter estimation for BNs are compelling [8, 9]. The Bayesian approach enables incorporation of expert opinion through the prior distribution and produces a posterior which aggregates the data and opinions. Further, the Bayesian approach enables calibration of some model variables with data even if other variables are not observed. We implemented a Bayesian approach to model calibration [10]. The calibration results were shown for a model describing Group Radicalization. 3.2 Data Driven Model Diagnostics Diagnostic checking is the act of inspecting the consistency of the modeling assumptions with the data. Inferences drawn from a calibrated model are only as trustable as the modeling assumptions are valid [11]. The major assumptions in specifying the structure of the BN are the conditional independencies among the variables. Thus, a methodology to inspect the validity of the conditional independencies in data is desirable. Statistical diagnostics are often approached through graphical inspection of model artifacts (such as residuals) for their consistency with the modeling assumptions. We accomplish this for BNs by plotting the hypothesized conditional independencies by using back-to-back bar charts. If the assumption (that is, the structure) is supported by the data, we should find reasonable symmetry in the back-toback distributions. The Group Radicalization model [10] assumes that the variables Isolated and Threatened and Competing for Support are conditionally independent, given the level of Radicalization. Figure 3 represents a visual inspection of this assumption in the

Bayesian Networks for Social Modeling

231

MAROB data [12]: if the assumption were true, the back-to-back distributions of Isolated and Threatened given each level of Competing for Support and Radicalization should be symmetric. The black dots represent a data driven quantification of this assumed symmetry. Their expectation is 0 under the hypothesized independence. The bands around the black dots given an uncertainty measure on this quantity are coded as green if the symmetry is reasonable, and red, otherwise. The graphic provides some evidence that, for the level of Radicalization in the "High" state, Isolated/Threatened is not independent of Competing for Support. Further, the graphic indicates that groups that are competing for supporters are less likely to be isolated and threatened than groups that are not competing. This result indicates that we may wish to alter the model structure by adding a direct dependence between Isolated and Threatened and Competing for Support in the BN.

Fig. 3. Bayesian network model fragment representing one aspect of radicalization

4 Evidence: Assessment and Linkage with BN Models After a model has been constructed and calibrated, the next step is typically to apply the model. A common analytic task is to assess the change in likelihood of the outcome of the model in light of observed evidence such as news reports, sensor data, and first hand testimonial. This is commonly accomplished in Bayesian network analysis by "setting" the state of an indicator to 1, and allowing the BN computation engine to compute the probabilities of the other nodes’ states given that new information. We have developed a more nuanced approach to evidence handling to reflect fundamental thinking in evidence assessment. Schum [13] distinguishes between an event and the evidence of the event, and identifies three attributes of credibility that must be considered when determining the inferential force that a piece of evidence has on the model: • Veracity – is the witness or attester of the event expected to report truthfully? • Observational sensitivity – is it possible for the witness to have observed the event? • Objectivity – does the witness of the event have a stake in the outcome? Likewise, our Bayesian network models assess the effect of whether or not an event occurs on the model separately from the likelihood of an event, given the evidence at

232

P. Whitney et al.

hand. Since the indicators in our social science models are taken from academic papers, they tend to be broad categories. The events, which are suggested by the model author or added by the user, are observable and concrete items. When the analyst adds a piece of evidence to the model, they are asked to choose the event to which it is relevant, whether it supports or refutes the event, and to rate the credibility and bias of the evidence. When evidence is attached to the Bayesian network model in the manner described above, the probabilities of the states of all indicators in the model are changed. Our system displays these changes for the user in a bar chart (Figure 4), showing the ratio of the probability with evidence over the probability before evidence. The X-axis of this plot is a quasi-log scale. The center tick means no change, the ticks to the right mean ratios of 2, 3 and 4, respectively (i.e. the probability doubled, tripled or quadrupled due to the evidence), and the ticks to the left mean 1/2, 1/3 and 1/4 (the probability was halved, etc, due to the evidence). Figure 4 shows that the probability of Proliferation Propensity being Very High was roughly halved due to the evidence that the user added to the model.

Fig. 4. Change in probabilities of proliferation propensity resulting from evidence

5 Example: Nuclear Non-proliferation All of the modeling pieces described above – a Bayesian network model, calibration with data and experts’ assessments, and the evidence gathering and attachment tool – form an integrated modeling framework, which we have applied to nuclear proliferation risk assessments. Below we discuss how this novel approach was employed to create case studies of ranking state-level proliferation propensity. 5.1 Model Development and Motivation The proliferation risk model [14] integrates technical, social, and geopolitical factors into a comprehensive modeling framework to estimate the state actors’ propensity to acquire nuclear weapons capability. This modeling approach benefits from insight offered by both technical proliferation modeling and social modeling, and is well grounded in academic literature. The Bayesian network model serves as the basic structure on which evidence is organized and relevant queries developed. We initially quantified the model with expert judgment, and subsequently with consolidated and enhanced data sets [14]. While these data were useful in defining base rates and have allowed us to exercise the model for specific countries, the conditional probability tables are still based upon expert judgment in various forms.

Bayesian Networks for Social Modeling

233

Fig. 5. Bayesian network model fragment for whether a country has a nuclear program

5.2 Data Gathering Process The integrated data gathering approach consists of four steps: target identification, evidence gathering, evaluation and attachment, and model updating. We focused on Africa since states in this continent are markedly diverse in terms of economic development, technical capabilities, regime stability, and governance. We created country proliferation propensity rankings for the following countries: South Africa, Nigeria, Uganda, Gabon, Senegal, and Ghana. During the evidence gathering stage, researchers developed two query sets for most of the nodes in the model, and then the query sets were merged to reduce redundancy and ensure better search term representation. For instance, for the node named "Nuclear Capability," we created a broad search query such as "(Country name) has nuclear capability." In the same vein, for the node, "Democracy," which represents a state’s degree of democracy, a query, "(Country name) is democratic," was created. In some cases, multiple queries were created to provide broader search coverage. The queries were run and candidate evidence retrieved. We evaluated the evidence in terms of its reliability and bias in accordance with the regular criteria. 5.3 BACH and Results The model development and data gathering were done in a software tool we created called BACH (Bayesian Analysis of Competing Hypotheses). BACH is a Javabased tool for Bayes net model creation, evidence search and assessment, and evaluation of the effect of evidence on a user’s hypotheses. BACH allows the user to create a model (as shown in Figure 5) and to document the model creation process including links and references. The tool also includes the ability to construct and save evidence queries (which at the moment search over the internet, but can be adapted to search databases or local networks) and supports evidence importation and credibility and bias assessment. When evidence is attached or the model is changed, BACH instantly updates the new hypothesis probabilities and the outcome effects (shown in Figure 4).

234

P. Whitney et al.

BACH updated the outcome probabilities for state proliferation propensity to reflect the attached evidence. The state propensity rankings are shown in Table 1 below. The likelihood of developing a nuclear weapons program from most likely to least likely is: South Africa, Nigeria, Ghana, Senegal, Gabon, and Uganda. The outcome change ratio measures the magnitude of propensity change resulting from the attached evidence for a given state and is calculated as [1- (base rate – updated calculation)/base rate]. The outcome node has five states: very low, low, moderate, high, and very high, and the results suggest that the attached evidence produced the greatest impact on the "very high" state for South Africa and Nigeria, the "very low" state for Ghana and Senegal, and the "low" state for Gabon and Uganda. Table 1. Proliferation Propensity Rankings for Six African Countries Propensity Ranking 1 2 3 4 5 6

Country South Africa Nigeria Ghana Senegal Gabon Uganda

Propensity (p) 0.42 0.35 0.18 0.16 0.12 0.09

Max Outcome Change Ratio 2.12 1.76 1.34 1.31 1.13 1.23

State with Max Changes Very High Very High Very Low Very Low Low Low

6 Conclusions We have presented a "cradle to grave" perspective on how BN models can be constructed, validated, and used in socio-behavioral modeling. The modeling construction process is suitable for a wide range of human and organizational modeling activities. We have found it suitable for steady-state or a "slice of time" modeling and analysis settings. The behavioral mechanisms are proposed in the literature, once represented in the form of a BN model, can be the subject of critical empirical evaluations. Further, a well-thought approach for evidence assessment is also available in this framework. This set of features contributes to the scientific utility of this category of modeling approaches. Acknowledgments. Christian Posse contributed to the early developments of the research direction and to some details of the research arc described in this paper.

References 1. Jensen, F., Nielsen, T.: Bayesian Networks and Decision Graphs, 2nd edn. Springer, New York (2007) 2. McCauley, C., Moskalenko, S.: Mechanisms of Political Radicalization: Pathways Toward Terrorism. Terrorism and Political Violence 20, 416–433 (2008) 3. Kahneman, D., Tversky, A.: Judgment under Uncertainty: Heuristics and Biases. Science 4157, 1124–1131 (1974) 4. Renooij, S.: Probability Elicitation for Belief Networks: Issues to Consider. Knowledge Engineering Review 16, 255–269 (2001)

Bayesian Networks for Social Modeling

235

5. Green, P., Krieger, A., Wind, Y.: Thirty Years of Conjoint Analysis: Reflections and Prospects. Interfaces 73, S56–S73 (2001) 6. Walsh, S., Dalton, A., White, A., Whitney, P.: Parameterizing Bayesian Network Representations of Social-Behavioral Models by Expert Elicitation: A Novel Application of Conjoint Analysis. In: Proceedings of Workshop on Current Issues in Predictive Approaches to Intelligent and Security Analytics (PAISA), pp. 227–232. IEEE Press, Los Alamitos (2010) 7. Cox, D., Hinkley, D.: Theoretical Statistics. Chapman & Hall, London (1974) 8. Heckerman, D., Geiger, D., Chickering, D.: Learning Bayesian Networks: The Combination of Knowledge and Statistical Data. Journal of Machine Learning 20, 197–243 (1995) 9. Riggelson, C.: Learning Parameters of Bayesian: Networks from Incomplete Data via Importance Sampling. International Journal of Approximate Reasoning 42, 69–83 (2005) 10. Whitney, P., Walsh, S.: Calibrating Bayesian Network Representations of SocialBehavioral Models. In: Chai, S.-K., Salerno, J.J., Mabry, P.L. (eds.) SBP 2010. LNCS, vol. 6007, pp. 338–345. Springer, Heidelberg (2010) 11. Cleveland, W.: Visualizing Data. Hobart Press, Summit (1993) 12. Minorities at Risk Organizational Behavior Dataset, http://www.cidcm.umd.edu/mar 13. Schum, D.: The Evidential Foundations of Probabilistic Reasoning. John Wiley & Sons, New York (1994) 14. Coles, G.A., Brothers, A.J., Gastelum, Z.N., Thompson, S.E.: Utility of Social Modeling for Proliferation Assessment: Preliminary Assessment. PNNL 18438 (2009)

A Psychological Model for Aggregating Judgments of Magnitude Edgar C. Merkle1 and Mark Steyvers2 1

2

Department of Psychology Wichita State University [email protected] Department of Cognitive Sciences University of California, Irvine [email protected]

Abstract. In this paper, we develop and illustrate a psychologicallymotivated model for aggregating judgments of magnitude across experts. The model assumes that experts’ judgments are perturbed from the truth by both systematic biases and random error, and it provides aggregated estimates that are implicitly based on the application of nonlinear weights to individual judgments. The model is also easily extended to situations where experts report multiple quantile judgments. We apply the model to expert judgments concerning flange leaks in a chemical plant, illustrating its use and comparing it to baseline measures. Keywords: Aggregation, magnitude judgment, expert judgment.

1

Introduction

Magnitude judgments and forecasts are often utilized to make decisions on matters such as national security, medical treatment, and the economy. Unlike probability judgments or Likert judgments, magnitude judgments are often unbounded and simply require the expert to report a number reflecting his or her “best guess.” This type of judgment is elicited, for example, when an expert is asked to forecast the year-end value of a stock or to estimate the number of individuals afflicted with H1N1 within a specific geographic area. Much work has focused on aggregating the judgments of multiple experts, which is advantageous because the aggregated estimate often outperforms the individuals [1]. A simple aggregation method, the linear opinion pool, involves a weighted average of individual expert judgments. If experts are equally weighted, the linear opinion pool neglects variability in: (1) the extent to which experts are knowledgeable, and (2) the extent to which experts can translate their knowledge into estimates. As a result, a variety of other methods for assigning weights and aggregating across judges have been proposed. These methods have most often been applied to probability judgments [2,3]. In the domain of probability judgments, the supra-Bayesian approach to aggregation [4,5] has received attention. This approach involves initial specification J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, pp. 236–243, 2011. c Springer-Verlag Berlin Heidelberg 2011 

Aggregating Magnitude Judgments

237

of one’s prior belief about the occurrence of some focal event. Experts’ judged probabilities are then used to update this prior via Bayes’ theorem. This method proves difficult to implement because one must also specify a distribution for the experts’ judgments conditional on the true probability of the focal event’s occurrence. As a result, others have focused on algorithms for weighting expert judgments. Kahn [6] derived weights for a logarithmic opinion pool (utilizing the geometric mean for aggregation), with weights being determined by prior beliefs, expert bias, and correlations between experts. This method requires the ground truth to be known for some items. Further, Cooke [7] developed a method to assign weights to experts based on their interval (i.e., quantile) judgments. Weights are determined both by the extent to which the intervals are close to the ground truth (for items where the truth is known) and by the width of the intervals. This method can only be fully utilized when the ground truth is known for some items, but it can also be extended to judgments of magnitude. Instead of explicitly weighting experts in an opinion pool, Batchelder, Romney, and colleagues (e.g., [8,9]) have developed a series of Cultural Consensus Theory models that aggregate categorical judgments. These models account for individual differences in expert knowledge and biases, and they implicitly weight the judgments to reflect these individual differences. The aggregated judgments are not obtained by averaging over expert judgments; instead, the aggregated judgments are estimated as parameters within the model. Additionally, these models can be used (and are intended for) situations where the ground truth is unknown for all items. In this paper, we develop a model for aggregating judgments of magnitude across experts. The model bears some relationships to Cultural Consensus Theory, but it is unique in that it is applied to magnitude judgments and can handle multiple judgments from a single expert for a given item. The model is also psychologically motivated, building off of the heuristic methods for psychological aggregation described by Yaniv [10] and others. Finally, the model can be estimated both when the ground truth is unknown for all items and when the ground truth is known for some items. In the following pages, we first describe the model in detail. We then illustrate its ability to aggregate expert judgments concerning leaks at a chemical plant. Finally, we describe general use of the model in practice and areas for future extensions.

2

Model

For a given expert, let Xjk represent a magnitude prediction of quantile qk for event j (in typical applications that we consider, q1 = .05; q2 = .5; q3 = .95). All predictions, then, may be represented as a three-dimensional array X, with the entry Xijk being expert i’s prediction of quantile qk for event j. The model is intended to aggregate across the i subscript, so that we have aggregated quantile estimates for each event.

238

E.C. Merkle and M. Steyvers

We assume an underlying, “true” uncertainty distribution for each event j, with individual experts’ uncertainty distributions differing systematically from the true distribution. Formally, let Vj be the true uncertainty distribution associated with event j. Then (1) Vj ∼ N(μ∗j , σj2∗ ), where normality is assumed for convenience but can be relaxed. For events with known outcomes, we can fix μ∗j at the outcome. Alternatively, when no outcomes are known, we can estimate all μ∗j . We assume that each expert’s uncertainty distribution differs systematically 2 from the true distribution. The mean (μij ) and variance (σij ) of expert i’s distribution for event j is given by: μij = αi + μ∗j , 2 σij

=

(2)

σj2∗ /Di ,

where the parameter αi reflects systematic biases of expert i and receives the hyperdistribution: (3) αi ∼ N(μα , φα ), i = 1, . . . , N. The parameter Di reflects expertise: larger values reflect less uncertainty. Di is restricted to lie in (0, 1) ∀ i because, by definition, expert uncertainty can never be less than the “true” uncertainty. We assume that expert i’s quantile estimates arise from his/her unique uncertainty distribution and are perturbed by response error: Xijk ∼ N(Φ−1 (qk )σij + μij , γi2 ),

(4)

where Φ−1 (·) is the inverse cumulative standard normal distribution function. The response error reflects the fact that the expert’s quantile estimates are likely to stray from the quantiles of his/her underlying normal distribution. We are estimating a Bayesian version of the above model via MCMC, so we need prior distributions for the parameters. These are given by Di ∼ Unif(0, 1) μα ∼ N(0, 4)

γi2 ∼ Unif(0, 1000) φα ∼ Unif(0, 1000)

μ∗j ∼ N(0, 1.5) . σj2∗ ∼ Gamma(.001, .001)

(5)

The above priors are minimally informative, though this is not obvious for the priors on μ∗j and μα . We further describe the rationale for these priors in the application. An additional advantage of the Bayesian approach involves the fact that we can easily gauge uncertainty in the estimated ground truth. In summary, the model developed here is unique in that: (1) the ground truth for each item is estimated via a model parameter, and (2) specific distributional assumptions are very flexible. Psychological aspects of the model include the facts that: (1) judges are allowed to have systematic response biases; and (2) judges are allowed to vary in expertise.

Aggregating Magnitude Judgments

3

239

Application: Flange Leaks

To illustrate the model, we use data on the causes of flange leaks at a chemical plant in the Netherlands [11,12]. Ten experts estimated the number of times in the past ten years that specific issues (called “items” below) caused a failure of flange connections. Issues included “changes in temperature caused by process cycles,” “aging of the gasket,” and “bolts improperly tightened.” For each issue, the experts reported 5%, 50%, and 95% quantiles of their uncertainty distributions. There were fourteen issues assessed, yielding a total of 10 × 14 × 3 = 420 observations. The ground truth (i.e., true number of times the issue led to a failure in the past ten years) was known for eight of the fourteen issues. 3.1

Implementation Details

The judgments elicited from experts varied greatly from item to item, reflecting the fact that some issues were much more frequent than others. Figure 1 illustrates this phenomenon in more detail. The figure contains 8 panels, each of which displays individual expert judgments in grey (labeled E1 through E10). The dots with horizontal lines reflect experts’ 5%, 50%, and 95% quantile judgments, and the horizontal line in each panel reflects the truth. The vast differences in magnitude can be observed by, e.g., comparing judgments for item 5 with those for item 6 (keeping in mind that the y-axis changes for each item). Further, focusing on item 5, we can observe extreme intervals that are far from the truth (for item 5, close to zero). Both the vast differences in judged magnitudes and the large outliers lead to poor model performance. To minimize these issues, we fit the model to standardized, log-transformed judgments for all items. These transformations make model estimation more stable, and the resulting model estimates can be “untransformed” to obtain aggregated estimates on the original scale. The transformed data also lead to default prior distributions on the μ∗j : because we are dealing with standardized data, a reasonable prior would be N(0, 1). Because experts tend to under- or overestimate some quantities, we have found it helpful to increase the prior variance to, say, 1.5. This also leads to the prior variance on μα : because we are dealing with standardized data, it would be quite surprising to observe a mean bias outside of (−2, 2). We used three implementations of the model to aggregate estimates in different ways. In the first implementation (M1), we fit the model to expert judgments for all 14 items, excluding the ground truth. This implementation reflects a situation where the truth is unknown and estimates must be aggregated. In the second implementation (M2), we used the same data but included the ground truth for the first four items. This reflects a situation where the truth unfolds over time or was already known to the analyst, so that experts can be “graded” on some items but not others. The third model implementation (M3) held out the ground truth for one item at a time. Thus, the estimated interval for the

240

E.C. Merkle and M. Steyvers

●

Item 3 100

Item 2 50

100

Item 1 ●

●

●

●

● ●

● ●

●

●

●

● ●

● ●

●

●

●

● ●

●

●

● ● ●

●

e3 e2

90

0 e5

e4

e7 e6

e9 e8

avg m2

e10 m1

● ●

●

●

●

● ●

●

●

●

e1

m3

e3 e2

e4

e6

● ● ●

●

●

e1

e3 e2

● ● ●

●

e5 e4

● ●

e7 e6

● ●

● ● ● ●

●

e9 e8

● ●

●

● ● ●

● ●

● ● ●

1500000

Estimates

● ●

●

e3 e2

●

e5 e4

● ●

● ● ●

e7 e6

●

● ●

●

e9 e8

●

●

●

80

m3

e3 e2

e5 e4

● ●

e5

e7 e6

e9 e8

● ●

●

●

●

●

●

e9 e8

avg m2

e10 m1

m3

Item 6 ●

●

●

●

● ●

● ●

●

● ●

●

● ● ●

●

avg m2

e10 m1

● ●

● ●

● ● ●

●

●

●

●

●

●

●

●

● ●

●

●

●

●

●

●

●

●

●

●

●

● ● ●

e1

m3

●

●

●

e4

e7

●

● ● ●

●

●

e6

● ●

● ●

●

●

e1

● ●

e3 e2

● ● ●

●

e3 e2

●

●

e5 e4

e7 e6

e9 e8

avg m2

e10 m1

m3

● ● ●

● ● ●

avg m2

e10 m1

m3

●

● ● ●

● ● ●

● ● ●

● ●

● ● ● ● ●

● ●

●

●

● ●

60

●

● ●

●

●

●

●

● ●

40

Estimates ● ● ●

●

●

●

20

● ●

● ●

● ●

80 e1

●

●

● ● ●

avg m2

●

● ●

● ●

100

1000

●

● ● ●

● ● ●

● ●

●

Item 8

600

Estimates

200 0

● ●

● ● ●

e10 m1

●

e1

●

●

e8

● ●

m3

●

●

●

●

●

●

●

●

●

avg m2

e10 m1

●

●

● ●

● ●

●

Item 7

● ● ●

● ● ●

e9

●

0

0

●

●

●

●

● ●

500000

40000 20000

Estimates

●

●

e7

●

●

●

● ● ●

● ● ●

●

Item 5

●

●

● ●

e5

Item 4

●

●

70

30

●

●

e1

●

●

●

●

40

● ●

●

● ●

●

30

● ●

●

●

●

60

● ●

● ●

●

20

40

●

●

● ●

50

●

Estimates

40 ●

●

Estimates

●

●

●

10

●

20

●

10

60

●

Estimates

●

●

20

Estimates

80

● ● ●

● ● ●

e1

e3 e2

e5 e4

e7 e6

e9 e8

avg m2

e10 m1

m3

Fig. 1. Magnitude estimates of ten experts and four aggregation methods for eight items. Vertical lines with points reflects 5%, 50%, and 95% quantile judgments; “e” labels represent experts, and the “avg” label represents the unweighted average aggregation. “M1” represents model-based aggregations with no ground truth, “M2” represents model-based aggregations with ground truth for the first four items, and “M3” represents model-based aggregations for all items except the one being estimated. The horizontal line in each panel represents the ground truth.

first item utilized the ground truth for items 2–8. The estimated interval for the second item utilized the ground truth for items 1 and 3–8, and so on. Finally, for comparison, we also aggregated the estimates via an unweighted average. 3.2

Main Results

Figure 1 contains the four aggregated intervals (black lines), with avg being unweighted averages and M1, M2, M3 being the model implementations. It can be seen that, as compared to the experts’ reported intervals, the aggregated intervals are consistently narrower and closer to the truth. The model-based

Aggregating Magnitude Judgments

241

aggregations also result in narrower intervals, as compared to the simple average across experts. Comparing the three aggregated intervals with one another, it can be seen that the (M2) intervals are centered at the true value for Items 1–4. However, for items where the ground truth was not known (items 5–8), the intervals are very similar for all three methods. This implies that the model did not benefit from “knowing” the ground truth on some items. We now focus on statistical summaries of Figure 1. Table 1 displays four statistics for each expert and for the aggregated intervals. The first is the number of intervals (out of eight) that covered the truth. The other four are ratio measures, comparing each expert/model to the model-based aggregations with no ground truth (M1). These measures include the average ratio of interval widths, defined as: 8 1  (Xij3 − Xij1 ) Width = , (6) j3 − X j1 ) 8 j=1 (X  are aggregated quantiles from M1. The next measure where i is fixed and the X is a standardized bias ratio, defined as:  ∗ ∗ j |Xij2 − μj |/μj Bias =  , (7) j2 − μ∗ |/μ∗ |X j

j

j

where μ∗j is the ground truth. The last measure is the average likelihood of the ground truth under each expert’s interval (assuming normality), relative to the sum of likelihoods under the expert’s interval and under the aggregated interval from M1: 8 φ((μ∗j − Xij2 )/σij ) 1 LR = , (8) j2 )/ 8 j=1 φ((μ∗j − Xij2 )/σij ) + φ((μ∗j − X σj ) where φ() is the standard normal density function. This is a measure of the relative likelihood of the truth for the expert’s intervals, as compared to the M1 intervals. For the width and bias measures described above, values less than one imply that the expert (or other aggregation) did better than M1 and values greater than one imply that the expert (or aggregation) did worse than M1. The likelihood measure, on the other hand, must lie in (0,1). Values less than .5 imply that the expert did worse than M1, and values greater than .5 imply the converse. Table 1 contains the statistics described above for each expert and for the aggregations. Statistics for M2 are not included because its estimated intervals are centered at the ground truth for the first four items (leading to inflated performance measures). Examining the individual expert judgments, the table shows that no expert beats M1 on all four measures. Expert 8 is closest: her intervals are wider than the model, but she covers more items, exhibits less bias, and exhibits greater truth likelihoods. Examining the aggregations, M1 beats the unweighted average on both interval width and bias. The two aggregation methods are similar on coverage and

242

E.C. Merkle and M. Steyvers

Table 1. Statistics for individual experts’ judgments and model-based, aggregated judgments. “Coverage” is number of intervals that cover the truth (out of 8), “Width” is ratio of of mean interval width to that of M1, “Bias” is ratio of mean standardized, absolute bias to that resulting from M1, and “LR” is the likelihood of the truth for each interval relative to M1. Expert Coverage Width Bias LR e1 3 2.8 1.5 0.28 e2 6 2.1 2.1 0.53 e3 3 1.3 2.3 0.19 e4 4 4.1 8.4 0.29 e5 3 1.9 2.3 0.45 e6 2 1.7 1.9 0.34 e7 4 1.2 1.7 0.53 e8 7 1.8 0.9 0.61 e9 1 1.0 1.7 0.39 e10 4 1.1 0.4 0.36 avg 4 1.9 2.2 0.55 M1 4 – – – M3 2 0.9 1.1 0.26

likelihood. Comparing M1 with M3, M3 is worse on coverage and likelihood ratio. These findings, which generally match the visual results in Figure 1, imply that having access to the ground truth for some items did not improve model estimates.

4

Conclusions

The psychological model for aggregating magnitude judgments described in this paper exhibits good statistical properties, including an ability to consistently outperform expert judgments and judgments from simple aggregation methods. Unexpectedly, the model was unable to benefit from the ground truth (though see [13] for a similar result in a different context). We view this to be related to the model assumption that experts consistently over- or underestimate the ground truth across items. This assumption is violated if, say, experts are very knowledgeable about some items but not others. If the assumption is violated, then the model cannot obtain good estimates of expert biases because the biases are changing across items. This leads to the disutility of the ground truth. A potential solution would involve the addition of “item saliency” parameters into the model, though we may quickly find that the model is overparameterized. While these and other extensions may be considered, the model developed here exhibits reasonable performance and has not been tailored to the content area in any way. It may generally be applied to domains where magnitude judgments are utilized, and it is very flexible on the number of quantile judgments elicited, the presence of ground truth, and the extent to which all experts judge all items. The model also utilizes interpretable parameters, allowing the analyst to maintain

Aggregating Magnitude Judgments

243

some control over the model’s behavior. This feature may be especially useful in situations where the costs of incorrect judgments are high. More generally, the model possesses a unique set of features that give it good potential for future application.

References 1. Surowiecki, J.: The wisdom of crowds. Anchor, New York (2005) 2. Clemen, R.T., Winkler, R.L.: Combining probability distributions from experts in risk analysis. Risk Analysis 19, 187–203 (1999) 3. O’Hagan, A., Buck, C.E., Daneshkhah, A., Eiser, J.R., Garthwaite, P.H., Jenkinson, D.J., Oakley, J.E., Rakow, T.: Uncertain judgements: Eliciting experts’ probabilities. Wiley, Hoboken (2006) 4. Morris, P.A.: Combining expert judgments: A Bayesian approach. Management Science 23, 679–693 (1977) 5. Genest, C., Zidek, J.V.: Combining probability distributions: A critique and annotated bibliography. Statistical Science 1, 114–148 (1986) 6. Kahn, J.M.: A generative Bayesian model for aggregating experts’ probabilities. In: Proceedings of the 20th Conference on Uncertainty in Artificial Intelligence, pp. 301–308 (2004) 7. Cooke, R.M.: Experts in uncertainty: Opinion and subjective probability in science. Oxford University Press, New York (1991) 8. Batchelder, W.H., Romney, A.K.: Test theory without an answer key. Psychometrika 53, 71–92 (1988) 9. Karabatsos, G., Batchelder, W.H.: Markov chain estimation for test theory without an answer key. Psychometrika 68, 373–389 (2003) 10. Yaniv, I.: Weighting and trimming: Heuristics for aggregating judgments under uncertainty. Organizational Behavior and Human Decision Processes 69, 237–249 (1997) 11. Cooke, R.M.: Experts in uncertainty: Opinion and subjective probability in science. Oxford, New York (1991) 12. Cooke, R.M., Goossens, L.L.H.J.: TU Delft expert judgment data base. Reliability engineering and system safety 93, 657–674 (2008) 13. Dani, V., Madani, O., Pennock, D., Sanghai, S., Galebach, B.: An empirical comparison of algorithms for aggregating expert predictions. In: Proceedings of the Twenty-Second Conference Annual Conference on Uncertainty in Artificial Intelligence (UAI 2006), Arlington, Virginia, pp. 106–113. AUAI Press (2006)

Speeding Up Network Layout and Centrality Measures for Social Computing Goals Puneet Sharma1,2, Udayan Khurana1,2, Ben Shneiderman1,2, Max Scharrenbroich3, and John Locke1 1

Computer Science Department Human-Computer Interaction Lab 3 Department of Mathematics University of Maryland College Park MD 20740 {puneet,udayan,ben}@cs.umd.edu, {john.b.locke,max.franz.s}@gmail.com 2

Abstract. This paper presents strategies for speeding up calculation of graph metrics and layout by exploiting the parallel architecture of modern day Graphics Processing Units (GPU), specifically Compute Unified Device Architecture (CUDA) by Nvidia. Graph centrality metrics like Eigenvector, Betweenness, Page Rank and layout algorithms like Fruchterman-Rheingold are essential components of Social Network Analysis (SNA). With the growth in adoption of SNA in different domains and increasing availability of huge networked datasets for analysis, social network analysts require faster tools that are also scalable. Our results, using NodeXL, show up to 802 times speedup for a Fruchterman-Rheingold graph layout and up to 17,972 times speedup for Eigenvector centrality metric calculations on a 240 core CUDA-capable GPU. Keywords: Social Computing, Social Network Analysis, CUDA.

1 Introduction The enormous growth of social media (e.g. email, threaded discussions, wikis, blogs, microblogs), social networking (e.g. Facebook, LinkedIn, Orkut), and user generated content (e.g. Flickr, YouTube, Digg, ePinions) presents attractive opportunities for social, cultural, and behavioral analysts with many motivations. For the first time in history, analysts have access to rapidly changing news stories, national political movements, or emerging trends in every social niche. While temporal trend analysis answers many questions, network analysis is more helpful in discovering relationships among people, sub-communities, organizations, terms, concepts, and even emotional states. Many tools have been developed to conduct Social Network Analysis (SNA), but our effort is tied to NodeXL1 [3], a network data analysis and visualization [8] plug-in for Microsoft Excel 2007 that provides a powerful and simple means to graph data 1

NodeXL: Network Overview, Discovery and Exploration for Excel http://nodexl.codeplex.com/

J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, pp. 244–251, 2011. © Springer-Verlag Berlin Heidelberg 2011

Speeding Up Network Layout and Centrality Measures for Social Computing Goals

245

contained in a spreadsheet. Data may be imported from an external tool and formatted to NodeXL specifications, or imported directly from the tool using one of the supported mechanisms (such as importing an email, Twitter, Flickr, YouTube, or other network). The program maps out vertices and edges using a variety of layout algorithms, and calculates important metrics on the data to identify nodes and relationships of interest. While NodeXL is capable of visualizing a vast amount of data, it may not be feasible to run its complex computation algorithms on larger datasets (like those found in Stanford’s SNAP library2) using hardware that is typical for a casual desktop user. And we believe that in order for data visualization tools to reach their full potential, it is important that they are accessible and fast for users using average desktop hardware. Nvidia’s CUDA3 technology provides a means to employ the dozens of processing cores present on modern video cards to perform computations typically meant for a CPU. We believe that CUDA is appropriate to improve the algorithms in NodeXL due to the abundance of Graphical Processing Units (GPUs) on mid-range desktops, as well as the parallelizable nature of data visualization. Using this technology, we hope to allow users to visualize and analyze previously computationally infeasible datasets. This paper has two contributions. The first contribution is to provide the strategies and directions to GPU-parallelize the existing layout algorithms and centralities measures on commodity hardware. And the second contribution to apply these ideas to widely available and accepted visualization tool, NodeXL. To our knowledge, ours is the first attempt to GPU-parallelize the existing layout algorithms and centrality measures on commodity hardware. In our modified NodeXL implementation, we targeted two computationally expensive data visualization procedures: the Fruchterman-Rheingold [1] force-directed layout algorithm, and the Eigenvector centrality node metric.

2 Computational Complexity of SNA Algorithms Social Network Analysis consists of three major areas - Network Visualization, Centrality Metric Calculation and Cluster detection. Popular techniques for visualization use force-directed algorithms to calculate a suitable layout for nodes and edges. Computationally, these algorithms are similar to n-body simulations done in physics. The two most popular layout algorithms are Harel-Koren [2] and Fruchterman-Rheingold [1]. The computational complexity of the latter can roughly be expressed as O(k(V2+E)) and the memory complexity as O(E+V), where k is the number of iterations needed before convergence, V is the number of vertices and E is the number of edges in the graph. Centrality algorithms calculate the relative importance of each node with respect to rest of the graph. Eigenvector centrality [6] measures the importance of a node by the measure of its connectivity to other "important" nodes in the graph. The process of finding it is similar to belief propagation and the algorithm is iterative with the time complexity O(k.E). 2

SNAP: Stanford Network Analysis Platform http://snap.stanford.edu/index.html 3 What is CUDA? http://www.nvidia.com/object/what_is_cuda_new.html

246

P. Sharma et al.

With the increase in the number of vertices and edges, the computation becomes super-linearly expensive and nowadays analysis of a very highly dense graph is a reasonable task. Therefore, speedup and scalability are key challenges to SNA. 2.1 Approaches for Speedup and Scalability To achieve greater speedup, two approaches that immediately come to mind are faster algorithms with better implementation, and secondly, more hardware to compute in parallel. If we believed that the best algorithms (possible, or the best known) are already in place, the second area is worth investigating. However, it is the nature of the algorithm which determines that whether it can be executed effectively in a parallel environment or not. If the graph can be easily partitioned into separate workable sets, then the algorithm is suitable for parallel computing, otherwise not. This is because the communication between nodes in such an environment is minimal, and so all the required data must be present locally. We divide the set of algorithms used in SNA into three different categories based on feasibility of graph partitioning: • Easy Partitioning - Algorithm requires only the information about locally connected nodes (neighbors) e.g. Vertex Degree Centrality, Eigenvector centrality • Relatively Hard Partitioning - Algorithm requires global graph knowledge, but local knowledge is dominant and efficient approximations can be made. e.g. Fruchterman-Rheingold • Hard Partitioning - Data partitioning is not an option. Global graph knowledge is required at each node. e.g. Betweenness centrality which requires calculation of All Pairs of Shortest Paths. Algorithms that fall under the first category can be made to scale indefinitely using a distributed systems like Hadoop [7] and GPGPUs. For the second category, there is a possibility of a working approximation algorithm, whereas nothing can be said about the third category of algorithms.

3 Layout Speedup 3.1 Layout Enhancement In NodeXL layout algorithms, the Fruchterman-Rheingold algorithm was chosen as a candidate for enhancement due to its popularity as an undirected graph layout algorithm in a wide range of visualization scenarios, as well as the parallelizable nature of its algorithm. The algorithm works by placing vertices randomly, then independently calculating the attractive and repulsive forces between nodes based on the connections between them specified in the Excel workbook. The total kinetic energy of the network is also summed up during the calculations to determine whether the algorithm has reached a threshold at which additional iterations of the algorithm are not required. The part of the algorithm that computes the repulsive forces is O(N2), indicating that it could greatly benefit from a performance enhancement (the rest of the algorithm is O(|V|) or O(|E|), where v and e are the vertices and edges of the graph,

Speeding Up Network Layout and Centrality Measures for Social Computing Goals

247

respectively). The algorithm is presented in pseudo-code and computationally intensive repulsive calculation portion of the algorithm is italicized in a longer version of this paper [9]. Due to the multiple nested loops contained in this algorithm, and the fact that the calculations of the forces on the nodes are not dependent on other nodes, we implemented the calculation of these forces in parallel using CUDA. In particular, the vector V was distributed across GPU processing cores. The resulting algorithm, Super Fruchterman-Rheingold, was then fit into the NodeXL framework as a selectable layout to allow users to recruit their GPU in the force calculation process. 3.2 Results for Layout Speedup We ran our modified NodeXL on the computer with CPU (3 GHz, Intel(R) Core(TM) 2 Duo) and GPU (GeForce GTX 285, 1476 MHz, 240 cores). This machine is located in the Human Computer Interaction Lab at the University of Maryland, and is commonly used by researchers to visualize massive network graphs. The results of our layout algorithm are shown in the table below. Each of the graph instances listed may be found in the Stanford SNAP library: Table 1. Results of layout speedup testing

Graph Instance Name

#Nodes

#Edges

Super F-R run time (ms)

CA-AstroPh

18,772

396,160

1,062.4

84,434.2

79x

cit-HepPh

34,546

421,578

1,078.0

343,643.0

318x

soc-Epinions1

75,879

508,837

1,890.5

1,520,924.6

804x

soc-Slashdot0811

77,360

905,468

2,515.5

1,578,283.1

625x

soc-Slashdot0902

82,168

948,464

2,671.7

1,780,947.2

666x

Simple F-R run time Speedup (ms)

The results of parallelizing this algorithm were actually better than we had expected. The speedup was highly variable depending on the graph tested, but the algorithm appears to perform better with larger datasets, most likely due to the overhead of spawning additional threads. Interestingly, the algorithm also tends to perform better when there are many vertices to process, but a fewer number of edges. This behavior is the result of reaping the benefits of many processing cores with a large number of vertices, while the calculation of the repulsive forces at each node still occurs in a sequential fashion. A graph with myriad nodes but no edges would take a trivial amount of time to execute on each processor, while a fully connected graph with few vertices would still require much sequential processing time. Regardless, the Super Fruchterman-Rheingold algorithm performs admirably even in

248

P. Sharma et al.

the worst case of our results, providing a 79x speedup in the CA-AstroPh set with a 1:21 vertex to edge ratio. The ability for a user to visualize their data within two seconds when it originally took 25 minutes is truly putting a wider set of data visualization into the hands of typical end users.

4 Metric Speedup 4.1 Metric Enhancement The purpose of data visualization is to discover attributes and nodes in a given dataset that are interesting or anomalous. Such metrics to determine interesting data include Betweenness centrality, Closeness centrality, Degree centrality, and Eigenvector centrality. Each metric has its merits in identifying interesting nodes, but we chose to enhance Eigenvector centrality, as it is frequently used to determine the “important” nodes in a network. Eigenvector centrality rates vertices based on their connections to other vertices which are deemed important (connecting many nodes or those that are “gatekeepers” to many nodes). By the Perron-Frobenius theorem we know that for a real valued square matrix (like an adjacency matrix) there exists a largest unique eigenvalue that is paired with an eigenvector having all positive entries. 4.2 Power Method An efficient method for finding the largest eigenvalue/eigenvector pair is the power iteration or power method. The power method is an iterative calculation that involves performing a matrix-vector multiply over and over until the change in the iterate (vector) falls below a user supplied threshold. We outline the power method algorithm with the pseudo-code in Section 4.2 in [9]. 4.3 Power Method on a GPU For our implementation we divided the computation between the host and the GPU. A kernel for matrix-vector multiply is launched on the GPU, followed by a device-tohost copy of the resulting vector for computation of the vector norm, then a host-todevice copy of the vector norm and finally a kernel for vector normalization is launched on the GPU. Figure 4 in [9] outlines the algorithm flow to allow execution on the GPU. 4.4 Results for Eigenvector Centrality Speedup We tested our GPU algorithm on eleven graph instances of up to 19k nodes and 1.7m edges. As a comparison, we used the serial version of the algorithm implemented in NodeXL with C#. Also, we used the same machine used in section 3.2. Table 2 shows the speedups we achieved by running the algorithm on the GPU. Speedups ranged between 102x and 17,972x. We investigated relationship between the problem instances and the speedups by examining how problem size, measured by the number of edges, affected the speedup.

Speeding Up Network Layout and Centrality Measures for Social Computing Goals

249

There is an increase in the speed up as the size of the graph increases. We believe that this is due to the effects of scale on the serial CPU implementation for graphs with bigger sizes, whereas the overhead of dividing the data for a CUDA-based calculation was more of a detriment for smaller datasets. Table 2. Tabular view of the speedups

#Nodes

#Edges

Time GPU(sec)

2,625

99,999

0.187

23.0

769

33,312

0.015

1.6

17,425

1,785,056

0.547

9,828.0

9,414

851,278

0.187

UNC FB Network Princeton FB Network

18,163

1,533,602

1.874

6,596

586,640

0.094

Saket Protein

5,492

40,332

0.109

SPI Human.pin Wiki Vote Gnutella P2P AstroPh Collab

8,776

35,820

0.078

7,115 10,874 18,772

103,689 39,994 396,160

0.266 0.172 0.344

Graph Name Movie Reviews CalTech Facebook Oklahoma Facebook Georgetown Facebook

Time CPU (sec)

1,320.0 13,492.0 495.0 107.0 263.0 137.0 681.7 2,218.7

Speedup

123x 102x 17972x 7040x 7196x 5280x 798x 3366x 516x 3966x 6454x

5 Future Work One facet of this work that we hope to discover more about in the future is the scalability of the CUDA Super Fruchterman-Rheingold layout algorithm. We were fortunate to have large datasets available from the SNAP library, but given the 666x speedup obtained in the largest dataset tested, which contains 82,163 nodes, we have yet to determine the limits of the algorithm. It is clear that given the restricted amount of memory on today’s GPUs (about 1 GB on high-end cards) that eventually memory will become a bottleneck, and we have yet to determine where that limit occurs. There are obviously interesting networks that contain millions of nodes, and we hope to be able to provide significant speedup for those situations as well. Unfortunately, the Windows Presentation Foundation code which paints a graph on the visualization window proved to be the weakest link in the visualization of multi-million node

250

P. Sharma et al.

graphs. On large enough graphs, the WPF module will generate an out-of-memory error or simply not terminate. Overcoming this WPF shortcoming would be the first step towards finding the bounds of the Super Fruchterman-Rheingold algorithm. Additionally, there are other algorithms within NodeXL that could benefit from GPU-based parallelization. The Harel-Koren Fast Multiscale algorithm [2] is another heavily-used layout algorithm in data visualization that is currently implemented sequentially. Finally, the speedup achieved using the eigenvector centrality metric also bring other vertex metrics to our attention. In particular, Closeness and Betweenness centrality are two measures that are used frequently in data visualization, and have yet to be implemented using CUDA.

6 Conclusions NodeXL is a popular data visualization tool used frequently by analysts on typical desktop hardware. However, datasets can be thousands or millions of nodes in size, and a typical desktop CPU does not contain enough processing cores to parallelize the various algorithms contained in NodeXL sufficiently to execute in a reasonable timeframe. Nvidia’s CUDA technology provides an intuitive computing interface for users to use the dozens of processing elements on a typical desktop GPU, and we applied the technology to two algorithms in NodeXL to show the benefit that the application can gain from these enhancements. The results obtained from running these enhancements were impressive, as we observed a 79x - 804x speedup in our implemented layout algorithm. Further, the 102x - 17,972x speedup gained by parallelizing the Eigenvector centrality metric across GPU cores indicates that when a CUDA-capable GPU is available, it should be used to perform centrality metrics. Such improvement in these execution times brings the visualization of data previously infeasible on a standard desktop into the hands of anyone with a CUDA-capable machine, which includes low-to-middle end GPUs today. While we have by no means maxed out the parallelization possibilities in the NodeXL codebase, we hope to have provided enough preliminary results to demonstrate the advantage that CUDA technology can bring to data visualization.

Acknowledgements We would like to acknowledge Dr. Amitabh Varshney and Dr. Alan Sussman for their guidance through the course of this project. We are grateful to Microsoft External Research for funding the NodeXL project at University of Maryland.

References 1. Fruchterman, T., Reingold, E.: Graph Drawing by Force-directed Placement. Software – Practice And Experience 21(11), 1129–1164 (1991) 2. Harel, D., Koren, Y.: A Fast Multi-Scale Algorithm for Drawing Large Graphs. Journal of Graph Algorithms and Applications 6(3), 179–202 (2002)

Speeding Up Network Layout and Centrality Measures for Social Computing Goals

251

3. Hansen, D., Shneiderman, B., Smith, M.: Analyzing Social Media Networks with NodeXL: Insights from a Connected World. Morgan Kaufmann, San Francisco 4. Perer, A., Shneiderman, B.: Integrating Statistics and Visualization: Case Studies of Gaining Clarity During Exploratory Data Analysis. In: ACM SIGCHI Conference on Human Factors in Computing Systems, pp. 265–274 (2008) 5. Harish, P., Narayanan, P.J.: Accelerating large graph algorithms on the GPU using CUDA. In: Proc. 14th International Conference on High Performance Computing, pp. 197–208 (2007) 6. Centrality – Wikipedia, http://en.wikipedia.org/wiki/Centrality 7. Hadoop, http://wiki.apache.org/hadoop/ 8. Social Network Analysis, http://en.wikipedia.org/wiki/Social_network#Social_network_ analysis 9. Sharma, P., Khurana, U., Shneiderman, B., Scharrenbroich, M., Locke, J.: Speeding up Network Layout and Centrality Measures with NodeXL and the Nvidia CUDA Technology. Technical report, Human Computer Interaction Lab, University of Maryland College Park (2010)

An Agent-Based Simulation for Investigating the Impact of Stereotypes on Task-Oriented Group Formation Mahsa Maghami and Gita Sukthankar Department of EECS University of Central Florida Orlando, FL [email protected],[email protected]

Abstract. In this paper, we introduce an agent-based simulation for investigating the impact of social factors on the formation and evolution of task-oriented groups. Task-oriented groups are created explicitly to perform a task, and all members derive benefits from task completion. However, even in cases when all group members act in a way that is locally optimal for task completion, social forces that have mild effects on choice of associates can have a measurable impact on task completion performance. In this paper, we show how our simulation can be used to model the impact of stereotypes on group formation. In our simulation, stereotypes are based on observable features, learned from prior experience, and only affect an agent’s link formation preferences. Even without assuming stereotypes affect the agents’ willingness or ability to complete tasks, the long-term modifications that stereotypes have on the agents’ social network impair the agents’ ability to form groups with sufficient diversity of skills, as compared to agents who form links randomly. An interesting finding is that this effect holds even in cases where stereotype preference and skill existence are completely uncorrelated. Keywords: Group formation, Multi-agent social simulations, Stereotypes.

1 Introduction Group membership influences many aspects of our lives, including our self-identities, activities, and associates; it affects not only what we do and who we do it with, but what we think of ourselves and the people around us. It can also give rise to stereotypic thinking in which group differences are magnified and importance of individual variations are discounted. Thinking categorically about people simplifies and streamlines the person perception process [12], facilitating information processing by allowing the perceiver to rely on previously stored knowledge in place of incoming information [11]. Whereas a variety of stereotypes are based on real group differences (e.g., cultural stereotypes about food preferences), stereotypes based on relatively enduring characteristics, such as race, religion, and gender, have an enormous potential for error [11] and can give rise to performance impairments [13]. We hypothesize that when stereotypes are formed independent of real group differences, it can result in negative effects for the collective system. However, studying the long-term effects of stereotypes can be difficult, especially to quantify the effects over a population rather than an individual. In this paper, J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, pp. 252–259, 2011. c Springer-Verlag Berlin Heidelberg 2011 

An Agent-Based Simulation for Investigating the Impact of Stereotypes

253

we describe an agent-based simulation for evaluating the impact of stereotypes on the performance of task-oriented groups. Understanding the role that stereotypes play in group formation can refine existing theory while providing insight into the efficacy of methods for combating the adverse effects of stereotypes on behavior and decisionmaking. We base our simulation on a model of multi-agent team formation [6] since taskoriented groups share many characteristics with teams, although lacking in shared training experience. In our simulation, the population of agents is connected by a social network that is locally updated over time by unoccupied agents depending on their preferences. Stereotypes are represented as an acquired preference model based on prior experience and observable agent features. In multi-agent environments, stereotypes have been used to provide faster modeling of other agents [4,5] and to bootstrap the trust evaluation of unknown agents [2]. In contrast, we examine the case of non-informative stereotypes; stereotypes affect the agents’ preferences for forming social attachments but do not affect the agents’ willingness or ability to cooperate with other agents.

2 Problem Statement To explore the impact of stereotype on group formation and network evolution, we have selected a simple multi-agent system model first introduced by Gaston and desJardins [6] and used in [7,8] to describe team formation. Since task-oriented groups are similar to teams, this is a reasonable method for modeling the task performance of group behavior on shared utility tasks in absence of stereotypes. In this model, there is a population of N agents represented by the set A = {a1 , . . . , aN }. Each agent can be considered as a unique node in the social network and the connection between the agents is modeled by an adjacency matrix E, where eij = 1 indicates an undirected edge between agent ai and aj and the degree of agent ai is defined as di = aj ⊆A eij . Each agent is assigned a single skill given by σi ∈ [1, σ] where σ is the number of available skills. Accomplishing each task requires a coalition of agents with the appropriate skills. Tasks are globally advertised for γ time steps at fixed intervals μ. Each task, Tk , has a size requirement, |Tk |, and a |Tk |-dimensional vector of required skills, RTk , which are selected uniformly from [1, σ]. When a coalition has formed with the full complement of skills required for the task, it takes α time steps for the group to complete the task. 2.1 Group Formation In the no stereotype case, we simply follow the group formation algorithm used to allocate agents to teams in [7]. During the team formation process, each agent, ai , can be in one of three states, si , UNCOMMITTED, COMMITTED, or ACTIVE. An agent in the UNCOMMITTED state has not been assigned to any task. An agent in the COMMITTED state has been assigned to a task but is still waiting for the enough agents with the right skills to join the group. Finally, an ACTIVE agent is currently working on a task with a complete group possessing the right complement of skills.

254

M. Maghami and G. Sukthankar

On each iteration, agents are updated in random order. UNCOMMITTED agents have the opportunity to adapt their local connectivity (with probability of Pi ) or can attempt to join a group. If a task currently has no other agents committed to it, an agent can initiate a group with a probability that is proportional to the number of its immediate UNCOMMITTED neighbors defined as follows:  aj ⊆A eij I(si , UNCOMMITTED)  IPi = , (1) aj ⊆A eij where I(x, y) = 1 when x = y and 0 otherwise. Agents are only eligible to join task-oriented groups in their local neighborhood, where this is at least one link between the agent and the group members. The algorithm used by an agent to initiate or join a group is presented in Algorithm 1. Algorithm 1. Group formation algorithm for all TK ⊆ T do if |Mk | = 0 and si = UNCOMMITTED then r ← U nif ormRandom([0, 1]) if r ≤ Pi then if ∃r ∈ RTk : r = σi then Mk ← Mk ∪ {ai } si ← COMMITTED end if end if else if ∃aj : eij = 1, aj ∈ Mk and si = UNCOMMITTED then if ∃r ∈ RTk : r = σi and r is unfilled then Mk ← Mk ∪ {ai } si ← COMMITTED end if end if end for

2.2 Network Adaptation To adapt the network structure, the agents modify their local connectivity based on the notion of preferential attachment [1]. Therefore, the probability of connecting to a given node is proportional to that node’s degree. As mentioned before, at each iteration the agent can opt to adapt its connectivity, with probability Pi . Modifying its local connectivity does not increase the degree of the initiating agent since the agent severs one of its existing connections at random and forms a new connection. To form a new connection, an agent considers the set of its neighbors’ neighbors designated as Ni2 = {am : eij = 1, ejm = 1, eim = 0, m = i}. The adapting agent, ai , selects a target agent, aj ⊆ Ni2 , to link to based on the following probability distribution: dj P (ai −→ aj ) =  , (2) al ⊆N 2 dl i

where d is the degree of agents.

An Agent-Based Simulation for Investigating the Impact of Stereotypes

255

The results in [7] and [8] show that this simple algorithm can be used to adapt a wide variety of random network topologies to produce networks that are efficient at information propagation and result in scale-free networks similar to those observed in human societies. For the non-stereotype group formation condition, we did not observe any differences between random attachment and preferential attachment.

3 Learning the Stereotype Model As noted in a review of the stereotype literature [11], stereotypes are beliefs about the members of a group according to their attributes and features. It has been shown that the stereotypes operate as a source of expectancies about what a group as a whole is like as well as what attributes individual group members are likely to possess [9]. Stereotype influences can be viewed as a judgment about the members of a specific group based on relatively enduring characteristics rather than their real characteristics. Here, we do not attempt to capture the rich cognitive notions of stereotype formation → − as it applies to humans, but rather view a stereotype as a function F : V −→ S, → − mapping a feature vector of agents, V , to a stereotypical impression of agents in forming friendships, S, which we will designate as the stereotype value judgment. This value represents the agents’ judgments on other groups and is based on observable features rather than skills or prior task performance. In most contexts, humans possess two types of information about others: 1) information about the individual’s attributes and 2) the person’s long-term membership in stereotyped groups [9]. Therefore, to learn the stereotype model, the simulation offers → − these two sources of information, V and its corresponding S which are related to the agents’ group membership, for a specific period of time. In our simulation, this initial period lasts for I time steps and helps the collaborating agents gain experience about the attributes of different groups of agents. Note that membership in these groups is permanent and not related to the agent’s history of participation in short-term task-oriented groups. In our work, we propose that each agent, ai , can use linear regression to build its own function, Fi , and to estimate the stereotype value of another agent, aj , according to the − → observable features Vj . After the initial period, I time steps, the estimated stereotype − → value of agent aj by agent ai will be calculated as Sˆij = Fi (Vj ). In our model, this stereotype value judgment affects the connection of agents during the network adaptation phase, as we will describe in the following section. 3.1 Network Adaptation with Stereotype Value Judgments In the stereotype case, the group formation algorithm is the same as described in Algorithm 1 but the network adaptation is based on the learned stereotype. If an agent decides to adapt its local network, again with probability Pi , it will do so based on its own stereotype model. To adapt the local connectivity network, each agent uses its learned model to make stereotype value judgment on other neighboring agents. This network adaptation process consists of selecting a link to sever and forming a new link.

256

M. Maghami and G. Sukthankar

Specifically, the agent ai first searches for its immediate neighbor that has the lowest stereotype value judgment, aj , and severs that link. The agent then searches for a new agent as a target for link formation. To form this link, it searches its immediate neighbors and the neighbors of neighbors. First the agent selects the neighbor with the highest stereotype value judgment, am , for a referral as this agent is likely to be a popular agent in its neighborhood. Then the adapting agent, ai , will establish a new connection with an , one of the most popular neighbors of am , assuming that it is not already connected. an = argak ∈A,eik =0 max Sˆik . Note that all of these selections are the result of the stereotype value judgment model that agent ai has about the other agents in its neighborhood.

4 Evaluation 4.1 Experimental Setup We conducted a set of simulation experiments to evaluate the effects of stereotype value judgments on group formation in a simulated society of agents. The parameters of the group formation model for all the runs are summarized in Table 1. In task generation, each task is created with a random number of components less than or equal to ten and a vector of uniformly-distributed skill requirements. To generate the agent society, each agent is assigned a specific skill, a feature vector, and a class label. The agents’ skills are randomly generated from available skills. Inspired by [2], four different long-lasting groups with different feature vector distributions are used as the basis for stereotype value judgments. Agents are assigned a six-dimensional feature vector, with each dimension representing an observable attribute, and a hidden stereotype value judgment drawn from Gaussian distribution assigned to the group. Table 1(b) shows the mean and standard deviations of the Gaussian distributions and the observable feature vector assigned to each group. The observable feature vectors are slightly noisy. To indicate the existence of an attribute, a random number is selected from distribution N (0.9, 0.05) to be close to 1 and to indicate the lack of an attribute this number is selected from distribution N (0.1, 0.05) to be close to zero. During the initial training period, agents are allowed to observe the hidden stereotype value judgment to learn the classifier. During the rest of the simulation these values are estimated by the agent’s learned model. In these experiments all the runs start with a random geometric graph (RGG) as the initial network topology among the agents. A RGG is generated by randomly locating all the agents in a unit square and connecting two agents if their distance is less than or equal to a specified threshold, d [3]. The random network we generated is a modified version of the RGG, proposed by [7]. In this version the d is selected as a minimal distance among the agents to guarantee that all the agents have at least one link to other agents. When the initial network is generated, the group formation will run for an initial period with no adaptation. During these initial steps, the agents can form groups and participate in task completion to gain some experiences about other agents. Results are based on the average of 10 different runs with a different initial network for each run.

An Agent-Based Simulation for Investigating the Impact of Stereotypes (a) Experimental parameters Parameter N σ γ α μ |T |

NIterations NInitial

Value 100 10 10 4 2 max 10 8000 2000

257

(b) Stereotype groups and feature vectors

Descriptions

Group

Total number of agents Total number of skills Task advertising time steps Agents active time Task interval Number of task required skills Number of iterations Number of initial iterations

Mean Value

G1 G2 G3 G4

0.9 0.6 0.4 0.3

StDev f1 f2 f3 f4 f5 f6 0.05 0.15 0.15 0.1

X

X X

X X X X X X

4.2 Results Performance. The performance of the system, like [7], is calculated as follows: TSuccessf ullyDone P erf ormance = , Ttotal

(3)

which is the proportion of successfully accomplished tasks over the total number of introduced tasks in the environment. Figure 1 shows the performance of system in the system with stereotypes and without stereotypes by iteration. 1

Performance

0.8 0.6 0.4

Degree Stereotype

0.2 0 2000

2500

3000

3500

4000

4500 5000 Iterations

5500

6000

6500

7000

Fig. 1. The performance of task-oriented groups (with and without stereotypes) vs. the iterations

The main effect of the stereotype is to move network toward a sparse network structure with a dramatic increase in isolate nodes. This drop in performance is even more pronounced with fewer total agents. Network Structure. Here, we examine the network structure to determine the evolution of the agent society. Figure 2 shows the Fiedler-embedding [10] of networks in the final connectivity network of N = 200 agents with and without stereotype value judgments. The degree-based strategy moves the structure toward being similar to a scale-free network whereas with stereotype value judgments the network becomes progressively more star-shaped.

258 220 200

220

Class1 Class 2 Class 3 Class 4

Class 1 Class 2 Class 3 Class 4

200 180

160

160

140

140

120

120

Y

Y

180

M. Maghami and G. Sukthankar

100

100

80

80

60

60

40

40

20 0

50

100

20 0

150

20

40

60

80

100

120

140

X

X

Fig. 2. Fiedler embedding of the final network structures in non-stereotype (left) and stereotype (right) based network evolution (N = 200)

Effects of Rapid Attachment Modification. Here we examine the effects of modifying the parameter P , the probability of updating the network, on the performance of the system, both with and without stereotypes. We varied this parameter from 0.1 to 0.9 with a step size 0.2. Figure 3 shows the performance modification during 5000 iterations in both strategies. As shown in the figure, the performance does not change significantly with P values before a certain threshold. After that threshold, the performance drops dramatically, as the agents spend more time updating the network than accomplishing tasks. This threshold is dependent on the total agents and number of skills required in the environment. In both strategies, the performance has dropped by P = 0.7 but in stereotype strategy the system drops at an earlier iteration since the information transmission efficiency of the network has been sabotaged by the stereotype-value judgments. Stereotype 1

0.9

0.9

0.8

0.8

0.7

0.7 Performance

Performance

No Stereotype 1

0.6 0.5 0.4 P = 0.1 P = 0.3 P = 0.5 P = 0.7 P = 0.9

0.3 0.2 0.1 0 2000

2500

3000

3500

4000

4500 5000 Iterations

5500

6000

6500

P = 0.1 P = 0.3 P = 0.5 P = 0.7 P = 0.9

0.6 0.5 0.4 0.3 0.2 0.1

7000

0 2000

2500

3000

3500

4000

4500 5000 Iterations

5500

6000

6500

7000

Fig. 3. The effect of the network adaptation probability, Pi

5 Conclusion and Future Work In this paper we introduced an agent-based simulation for examining the effects of stereotypes on task-oriented group formation and network evolution. We demonstrate that stereotype value judgments can have a negative impact on task performance, even in the mild case when the agents’ willingness and ability to cooperate is not impaired. By modifying the social network from which groups are formed in a systematically suboptimal way, the stereotype-driven agents eliminate the skill diversity required for

An Agent-Based Simulation for Investigating the Impact of Stereotypes

259

successful groups by driving the network toward specific topological configurations that are ill-suited for the task. The results show that making connections with agents solely based on group membership yields a sparser network with many isolated nodes. In future work, we plan to explore the effects of different stereotype models on network topology. Additionally, we plan to incorporate EPA (Evaluation, Potency, and Accuracy) [14] style parameters into our stereotype judgment model. Due to the technical challenges of investigating the long-term effects of stereotype across populations, we suggest our agent-based simulation method is a useful tool for investigating these research questions.

Acknowledgments This research was funded by AFOSR YIP award FA9550-09-1-0525.

References 1. Albert, R., Barab´asi, A.: Statistical mechanics of complex networks. Reviews of Modern Physics 74(1), 47–97 (2002) 2. Burnett, C., Norman, T., Sycara, K.: Bootstrapping trust evaluations through stereotypes. In: Proceedings of the 9th International Conference on Autonomous Agents and Multiagent Systems, pp. 241–248 (2010) 3. Dall, J., Christensen, M.: Random geometric graphs. Physical Review E 66(1), 16121 (2002) 4. Denzinger, J., Hamdan, J.: Improving modeling of other agents using tentative stereotypes and compactification of observations. In: Proceedings of Intelligent Agent Technology, pp. 106–112 (2005) 5. Denzinger, J., Hamdan, J.: Improving observation-based modeling of other agents using tentative stereotyping and compactification through kd-tree structuring. Web Intelligence and Agent Systems 4(3), 255–270 (2006) 6. Gaston, M., Simmons, J., desJardins, M.: Adapting network structure for efficient team formation. In: Proceedings of the AAAI 2004 Fall Symposium on Artificial Multi-agent Learning (2004) 7. Gaston, M., desJardins, M.: Agent-organized networks for dynamic team formation. In: Proceedings of the International Conference on Autonomous agents and Multiagent Systems, pp. 230–237 (2005) 8. Glinton, R., Sycara, K., Scerri, P.: Agent-organized networks redux. In: Proceedings of AAAI, pp. 83–88 (2008) 9. Hamilton, D., Sherman, S., Ruvolo, C.: Stereotype-based expectancies: Effects on information processing and social behavior. Journal of Social Issues 46(2), 35–60 (1990) 10. Hendrichson, B.: Latent semantic analysis and Fiedler embeddings. In: Proceedings of the Fourth Workshop on Text Mining of the Sixth SIAM International Conference on Data Mining (2006) 11. Hilton, J., Von Hippel, W.: Stereotypes. Annual Review of Psychology 47 (1996) 12. Macrae, C., Bodenhausen, G.: Social cognition: Thinking categorically about others. Annual Review of Psychology 51(1), 93–120 (2000) 13. Marx, D.M.: On the role of group membership in stereotype-based performance effects. Journal of Social and Personality Psychology Compass 3(1), 77–93 (2009) 14. Osgood, C.: The measurement of meaning. Univ. of Illinois, Urbana (1975)

An Approach for Dynamic Optimization of Prevention Program Implementation in Stochastic Environments Yuncheol Kang and Vittal Prabhu Department of Industrial and Manufacturing Engineering, Pennsylvania State University, University Park, PA 16802, USA

Abstract. The science of preventing youth problems has significantly advanced in developing evidence-based prevention program (EBP) by using randomized clinical trials. Effective EBP can reduce delinquency, aggression, violence, bullying and substance abuse among youth. Unfortunately the outcomes of EBP implemented in natural settings usually tend to be lower than in clinical trials, which has motivated the need to study EBP implementations. In this paper we propose to model EBP implementations in natural settings as stochastic dynamic processes. Specifically, we propose Markov Decision Process (MDP) for modeling and dynamic optimization of such EBP implementations. We illustrate these concepts using simple numerical examples and discuss potential challenges in using such approaches in practice. Keywords: Markov Decision Process, Partially Observable Markov Decision Process, Evidence-based Prevention Program.

1

Introduction

The science of preventing youth problems has significantly advanced in developing evidence-based prevention program (EBP) by using randomized clinical trials [1]. Effective EBP can reduce delinquency, aggression, violence, bullying and substance abuse among youth. For instance, studies have shown the annual cost of substance abuse in the U.S. to be $510.8 billion in 1999, and effective school-based prevention programs could save $18 for every $1 spent on these programs [2]. Some of the widely used EBPs include PATHS (Promoting alternative Thinking Strategies), BBBS (Big Brothers Big Sisters), SFP (Strengthening Families Program) and LST (LifeSkills Training). EBP services are commonly delivered to youth, and sometimes their families, individually or in groups through school and community settings, i.e., the natural world settings. Usually the outcomes of EBP implemented in natural settings tend to be lower than in the clinical trials [3-5]. One of the main challenges in improving the outcomes of EBP implementations is to develop a better understanding of the complex interplay between EBP and the natural world by identifying the key factors that influence the outcomes and the manner in which these factors and the interplay evolve during an implementation [3]. Recently adaptive intervention dosage control in EBP has been proposed as a potential method to improve the outcomes while reducing the cost [6-7]. Sustaining effective EBP services under unknown and J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, pp. 260–267, 2011. © Springer-Verlag Berlin Heidelberg 2011

Dynamic Optimization of Prevention Program Implementation

261

uncertain dynamics with significant budget and capacity constraints presents an intellectual challenge and is an acute social imperative. In this article, we present an approach for optimizing EBP implementations in stochastic dynamic environments. Our presentation is especially focused on the Markov Decision Process (MDP) approach for dynamically optimizing EBP implementations. We review MDP approach, and illustrate its potential application using numerical examples. Lastly, we discuss the challenges in applying the MDP approach for EBP implementations in practice.

2

Dynamic Optimization in Stochastic Environments

Many situations require a series of decisions to be made sequentially over time to optimize some overall objective. These problems are characterized by the need to consider the objective over the entire time horizon rather than the immediate reward at any one time step. A wide range of such problems in engineering and economics have been solved using dynamic optimization approaches since the pioneering work of Bellman in the 1950s [8]. In many situations, changes in the external and exogenous environment randomly influences the outcomes, and the decisions made at various time steps has some influence on the outcomes. Such stochastic dynamic optimization problems can be modeled as Markov Decision Processes (MDP). MDP models have been widely used in a variety of domains including management, computer systems, economics, communication, and medical treatment [9]. The “randomness” in MDP models can be caused by incomplete understanding of the process itself, or by uncertainties in interactions between the process and the environment. MDP is a variant and extended type of Markov Chain, which is a stochastic process having the property such that the next state depends on only the current state. MDP modeling approach is best suited for situations in which there are multiple actions which we can control over a Markov chain. To construct an MDP model, we need to define the following: state, action, decision epoch, and reward. First, a state is a random variable that represents the evolution of the system at any given time. As the system evolves it transitions from one state to another. Meanwhile, an action indicates a possible decision in the system, which in turn influences the current state, and transition to another state. A decision epoch refers to the discrete time when we take an action, and we need to determine the length of the decision horizon. The solution to an MDP problem enables us to choose a set of actions, also known as a policy, which maximizes the targeted value, usually called a reward. In other words, a reward occurs when an action is taken at the current state, and the algorithm for solving MDP results in the policy maximizing reward over the decision horizon. After an MDP model is constructed, we can obtain an optimal policy from the optimality equation, and corresponding structural results of the dynamics in the model. For the optimal policy, a number of algorithms have been introduced such as value iteration, policy iteration or linear programming [9]. Several approaches also have been developed to reduce computational burden in solving an MDP problem [10].

262

2.1

Y. Kang and V. Prabhu

Modeling EBP Implementation as MDP

In this paper we propose that EBP implementations can be modeled as dynamic optimization problems in which a sequence of decisions are made over multiple time periods through which the best possible objective is obtained. More particularly, we propose that EBP implementations can be modeled as an MDP. For illustration, let us consider one of the widely used EBP, PATHS (Promoting Alternative THinking Strategies). PATHS is a school-based program for promoting emotional and social competencies and reducing aggression and behavior problems in elementary schoolaged children while simultaneously enhancing the educational process in the classroom [11]. PATHS is a “universal” program in the sense it is delivered to all the youth in the class. PATHS curriculum provides teachers with systematic, developmentally-based lessons, materials, and instructions. There are two major stochastic processes in PATHS implementations. The first major stochastic process is various dynamic transitions in a youth’s mental state. We regard this as a stochastic process from a modeling perspective because the actual learning process in a specific youth is too complex, and current science cannot deterministically predict the social and emotional changes in the youth caused by PATHS. The second major stochastic process is the dynamically changing environment around the youth. The net combined effect of PATHS and the environment is therefore modeled as causing random transitions in a youth’s mental state, . A crucial aspect of the model is to quantify the probability of transitioning from one state to another. A state can be defined as a scalar or vector quantity depending on the intended purpose and richness of the model. For example, in a PATHS implementation, a youth’s Social Health Profile (SHP) such as authority acceptance, cognitive concentration and social competence can be used to represent the state vector for the youth, which can be measured through periodic observations by teacher(s) and others [12]. A state could also include a youth’s aggressive, hyperactive-disruptive, and prosocial behaviors measured through peer reports [12]. Depending on the purpose of the model, a state can be vector or a scalar index that quantifies the youth’s social and emotional state. Instead of modeling at the individual youth level we could also consider at a higher level of aggregation such as a class room or school. The set of all the states and the transition probabilities can be used to derive statistical properties of the system’s future. Possible actions, , for controlling PATHS implementation could be improving the dosage and fidelity of the implementation, providing additional technical assistance to the teacher(s) delivering the program, or providing additional intensive intervention to high risk youth. The premise of MDP modeling here is that the actions taken will influence the dynamic state transitions even though the actions themselves cannot completely determine the outcome. The menu of actions that are available will depend on the available budget and resources for the specific implementation in practice. A series of such actions are taken at each decision epoch, t, over the decision horizon, . Decision epoch in PATHS could be every month or every academic semester, and can vary depending on the specific implementation. The decision horizon can be an academic year or multiple years depending on the planning horizon , , of the particular implementation. There is a quantitative reward, corresponding to each state and action. In EBP implementations reward can be a

Dynamic Optimization of Prevention Program Implementation

263

combination of costs and benefits for the state and action. In particular, the cost can be regarded as a negative reward while the benefit can be treated as a positive reward. , to maximize the reward over The core problem of MDP is to select the policy, the planning horizon by solving the following optimality equation. | ,

,

. for

| , where is the state at 1, when the action is taken at and 2.2

1, … ,

is the transition probability from is the expected future reward.

1 (1) to

Advanced Modeling of Stochastic Processes in EBP Implementations

As prevention science unravels the “black box” model of the youth’s mental processes we would be able to refine and improve the stochastic process models proposed above. A key issue that is likely to remain with us is that the youth’s mental state will not be directly accessible. Any observations by teachers, parents, peers or self will at best be a probabilistic estimate of the true mental state. We can model such situations using a special case of MDP approach called Partially Observable Markov Decision Process (POMDP). In POMDP, the actual states are considered as hidden and are not directly accessible to observers. Rather, the observers have a set of observable outputs, , probabilistically related to the actual states and a set of belief, , based on the observations [13]. The belief is updated using a belief update | , , . Thus, POMDP selects the policy, function, , to maximize the reward based upon beliefs of the states by solving the following equation. ,

| ,

| , , for

. 1, … ,

1 (2)

| , is the where and is the observation and the belief at 1, probability of next observation based on the current belief and action. Since the current action influences the actual state transition, the probability of next observation is implicitly based on the state transition caused by current action chosen. | , , is the expected future reward. POMDP approaches have been used for modeling diseases that cannot be directly observed in patient’s symptoms [14-15]. EBP implementations could be modeled as POMDP in which teachers observation report is used for probabilistic estimation of the youth’s true mental state using belief functions. The belief function in POMDP would be an attractive mechanism to model the dynamics of a youth’s social and emotional development trajectory for a given combination of EBP and social environment. While POMDP offers a more advanced modeling approach to deal with stochastic dynamics of EBP implementations, they are computationally much harder to solve compared to MDP.

264

3

Y. Kang and V. Prabhu

Illustration of MDP Model for EBP Implementation

In this section we illustrate stochastic models for EBP implementations using two simple numerical examples. The first example illustrates the stochastic dynamics of EBP and the second example indicates how such dynamics can be potentially influenced by decision making using a MDP model. It should be emphasized that these examples are not based on real data. 3.1

Illustration 1: Stochastic Dynamics of EBP

Let us consider the Social Competence (SC) measure in SHP as the state for this illustration. Fig. 1 graphically represents the Markov Chain that models the stochastic dynamics using state transition probabilities without any prevention intervention. Fig. 2 represents the Markov Chain model with prevention intervention. These stochastic models can be used for a variety of analysis such as predicting the long-run social competence by computing the steady state probabilities, which is illustrated in Fig. 3. We can also use this to readily estimate long-run reward by using cost and/or benefit of each state.

Fig. 1. Dynamics of SC without prevention intervention. Each circle represents the state indicating the level of SC of the individual participant. The level of SC in Fig. 1 and 2 is abbreviated as Lv. The higher level of SC represents the better outcome in terms of social competence. Arcs are the transition between the states, and the number on the arc is the probability of the transition.

Fig. 2. Dynamics of SC with prevention intervention. Note that in this model there is higher probability of transitioning to a better SC state and compared to Fig. 1.

Dynam mic Optimization of Prevention Program Implementation

3.2

265

Illustration 2: Decission Makings in EBP Implementation Using MDP

The Markov Chain models in the previous illustration can be extended for optimizzing mplementations using an MDP model. To illustrate this we decision making in EBP im use the transition probabilitties shown in Fig. 2. Additionally, suppose the cost of the prevention intervention is 7 units for each decision epoch, and we assign 0, 13, 15, 18, 20 units as the amount of benefit b from SC level 1 to level 5, respectively. We deffine reward for each pair off state and action, and calculate it by summing the nefit. Since MDP chooses the best action for each decission corresponding cost and ben epoch, we can make the besst decision in terms of maximizing cumulative reward frrom the current decision epoch onwards. o

Fig. 3. Steady State Probaability for scenarios with and without prevention intervention

Fig. 4. Optimal Action n in SC level 1(left panel) and SC level 5 (right panel)

Fig. 4 shows two differeent scenarios of decision makings for the lowest SC level, SC level 1(left panel) and for f the highest SC level, SC level 5 (right panel). The left panel shows prevention inttervention results in a higher reward if it is delivered for

266

Y. Kang and V. Prabhu

more than two decision epochs in case of the lowest SC level. On the contrary, if the current state is in the highest SC level, then without prevention is always more beneficial than the other in terms of reward (right panel in Fig. 4). In many situations in EBP implementation, one often faces with considering a trade-off between the benefit and cost of prevention intervention. Under the circumstance, an MDP model would provide a guideline for indicating whether the cost of prevention intervention is enough to justify the benefit as this example describes.

4

Challenges in Applying MDP Approach

MDP modeling of prevention implementation could help researchers and practitioners obtain new insights into underlying dynamics and cost-effectiveness of prevention interventions. However, there are several challenges in using the MDP approach for EBP. The first challenge is to identify the structure and parameters of the MDP model such as the appropriate definition of the state and estimation of state transition probabilities. In particular, state transition probabilities and reward should reflect the characteristics of the EBP implementation with adequate accuracy. A related issue is to assign transition probabilities from sources such as prior randomized clinical trials of EBP. It should be noted that the MDP approach provides the optimal policy which maximizes the reward consisting of cost and benefit. Therefore the usefulness of the approach will be limited by accuracy of the cost and benefit estimates, especially in any monetization. Also, POMDP models can become computationally intractable even for modest sized problems and are likely to require carefully crafted solution techniques especially suitable for EBP implementation [9-10].

5

Conclusion

The need to improve the effectiveness of EBP in natural world settings has motivated the need to study EBP implementations. There is a dearth of models that mathematically characterize the dynamics of EBP processes at the individual or group levels - there is a pressing need to unravel this black box [16]. In this paper we proposed Markov Decision Process (MDP) for modeling and dynamic optimization of such EBP implementations. We illustrated these concepts using simple numerical examples and discussed potential challenges in using such approaches in practice. The modeling approaches proposed in this paper could help advance Type II translational research to create a new science of adoption and implementation of EBP in the real world and enable their ultimate goal of wide-spread public health impact [5].

Acknowledgements We gratefully acknowledge numerous discussions with Meg Small and Celene Domitrovich of Penn State’s Prevention Research Center over the last several years that motivated this work.

Dynamic Optimization of Prevention Program Implementation

267

References 1. Greenberg, M., Domitrovich, C., Bumbarger, B.: Preventing mental disorders in schoolage children: A review of the effectiveness of prevention programs. Prevention Research Center for the Promotion of Human Development, Pennsylvania State University. Submitted to Center for Mental Health Services, Substance Abuse and Mental Health Services Administration, US Department of Health & Human Services (1999) 2. Miller, T., Hendrie, D.: Substance abuse prevention dollars and cents: A cost-benefit analysis. In: Services, US Department of Health & Human Services. (ed.), p. 59 (2009) 3. Greenberg, M.: Current and future challenges in school-based prevention: The researcher perspective. Prevention Science 5, 5–13 (2004) 4. Bumbarger, B., Perkins, D.: After randomised trials: issues related to dissemination of evidence-based interventions. Journal of Children’s Services 3, 55–64 (2008) 5. Bumbarger, B.K., Perkins, D.F., Greenberg, M.T.: Taking effective prevention to scale. In: Doll, B., Pfohl, W., Yoon, J. (eds.) Handbook of Prevention Science. Routledge, New York (2009) 6. Collins, L., Murphy, S., Bierman, K.: A conceptual framework for adaptive preventive interventions. Prevention Science 5, 185–196 (2004) 7. Rivera, D., Pew, M., Collins, L.: Using engineering control principles to inform the design of adaptive interventions: A conceptual introduction. Drug and Alcohol Dependence 88, S31–S40 (2007) 8. Bellman, R., Lee, E.: History and development of dynamic programming. Control Systems Magazine 4, 24–28 (2002) 9. Puterman, M.: Markov decision processes: Discrete stochastic dynamic programming. John Wiley & Sons, Inc., New York (1994) 10. Schaefer, A., Bailey, M., Shechter, S., Roberts, M.: Modeling medical treatment using Markov decision processes. Operations Research and Health Care, 593–612 (2005) 11. Greenberg, M., Kusche, C., Cook, E., Quamma, J.: Promoting emotional competence in school-aged children: The effects of the PATHS curriculum. Development and Psychopathology 7, 117–136 (1995) 12. Bierman, K., Coie, J., Dodge, K., Greenberg, M., Lochman, J., McMahon, R., Pinderhughes, E.: The Effects of a Multiyear Universal Social-Emotional Learning Program: The Role of Student and School Characteristics. Journal of Consulting and Clinical Psychology 78, 156–168 (2010) 13. Smallwood, R., Sondik, E.: The optimal control of partially observable Markov processes over a finite horizon. Operations Research 21, 1071–1088 (1973) 14. Hauskrecht, M., Fraser, H.: Planning treatment of ischemic heart disease with partially observable Markov decision processes. Artificial Intelligence in Medicine 18, 221–244 (2000) 15. Kirkizlar, E., Faissol, D., Griffin, P., Swann, J.: Timing of testing and treatment for asymptomatic diseases. Mathematical Biosciences, 28–37 (2010) 16. Oakley, A., Strange, V., Bonell, C., Allen, E., Stephenson, J.: Process evaluation in randomised controlled trials of complex interventions. Bmj 332, 413 (2006)

Ranking Information in Networks Tina Eliassi-Rad1 and Keith Henderson2

2

1 Rutgers University [email protected] Lawrence Livermore National Laboratory [email protected]

Abstract. Given a network, we are interested in ranking sets of nodes that score highest on user-specified criteria. For instance in graphs from bibliographic data (e.g. PubMed), we would like to discover sets of authors with expertise in a wide range of disciplines. We present this ranking task as a Top-K problem; utilize fixed-memory heuristic search; and present performance of both the serial and distributed search algorithms on synthetic and real-world data sets. Keywords: Ranking, social networks, heuristic search, distributed search.

1 Introduction Given a graph, we would like to rank sets of nodes that score highest on user-specified criteria. Examples include (1) finding sets of patents which, if removed, would have the greatest effect on the patent citation network; (2) identifying small sets of IP addresses which taken together account for a significant portion of a typical day’s traffic; and (3) detecting sets of countries whose vote agree with a given country (e.g., USA) on a wide range of UN resolutions. We present this ranking task as a Top-K problem. We assume the criteria is defined by L real-valued features on nodes. This gives us a matrix F , whose rows represent the N nodes and whose columns represent the L realvalued features. Examples of these node-centric features include degree, betweenness, PageRank, etc. We define an L-tuple v1 , . . . , vL  as a selection of one value from each column. The score of an L-tuple is equal to the sum of the selected values (i.e., sum of the tuple’s components). For a given parameter K, we require an algorithm that efficiently lists the top K L-tuples ordered from best (highest score) to worst. The Ltuple with the highest score corresponds to a set of nodes in which all features take on optimal values. As described in the next section, we solve the Top-K problem by utilizing SMA* [4], which is a fixed-memory heuristic search algorithm. SMA* requires the specification of an additional parameter M for the maximum allotted memory-size. The choice for M has a dramatic effect on runtime. To solve this inefficiency problem, we introduce a parallelization of SMA* (on distributed-memory machines) that increases the effective memory size and produces super-linear speedups. This allows use to efficiently solve the Top-K ranking problem for large L and K (e.g., L ∈ [102 , 103 ] and K ∈ [106 , 109 ]). Experiments on synthetic and real data illustrate the effectiveness of our solution. J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, pp. 268–275, 2011. c Springer-Verlag Berlin Heidelberg 2011 

Ranking Information in Networks

269

2 A Serial Solution to the Top-K Ranking Problem Given F , a matrix of L columns of real-valued features over N rows of vertices, and an integer K, our application of SMA* will report the top K scoring L-tuples that can be selected from F . First, we take each column of matrix F and represent it as its own vector. This produces a list X containing L vectors (of size N × 1). We sort each vector in X in decreasing order. Then, we begin constructing D, another list of L vectors of real numbers. Each vector Di has |Xi | − 1 entries, where Dij = Xij − Xi(j+1) . Note that the values in each Di are all nonnegative real numbers; and Di is unsorted. The top-scoring L-tuple, R1 , can immediately be reported; it is simply the tuple generated by selecting the first element in each Xi . At this point, the problem can be divided into two subproblems whose structure is identical to the original problem (i.e. another instance of the Top-K problem). While there are several possible ways to make this division, we choose one that allows us to search efficiently. In particular, we choose the vector which would incur the least cost when its best element is discarded.1 This can be determined by choosing the vector i with the minimum Di1 . Given this index i, we generate two subproblems as follows. In the first subproblem, we discard the best element in vector Xi . The resulting list of vectors will be called X 1 . In the second subproblem, we keep the best element in vector Xi and remove all other elements from vector Xi . The resulting list of vectors here will be called X 0 . Let’s illustrate this procedure with an example. Suppose X = {[10,8,5,2,1], [4,3,2,1], [30,25,24,23,22]}; so, D = {[2,3,3,1], [1,1,1], [5,1,1,1]}. (Recall that we assume lists in X are already sorted in decreasing order and some entries maybe missing.) The topscoring L-tuple is R1 = 10, 4, 30 with score(R1 )=10+4+30=44. Starting from X, the next best tuple can be generated by selecting the list i with the smallest Di1 and decrementing Xi in X: X 1 = {[10,8,5,2,1], [3,2,1], [30,25,24,23,22]}, D1 = {[2,3,3,1], [1,1], [5,1,1,1]}, and score(X 1 )=10+3+30=43. At this point, we can split the problem into two smaller problems. We can either “accept” the best decrement (X 1 above) or “reject” the best decrement and all future chances to decrement that list: X 0 = {[10,8,5,2,1], [4], [30,25,24,23,22]}, D0 = {[2,3,3,1], [], [5,1,1,1]}, and score(X 0 )=10+4+30=44. In this way, we can generate X 11 , X 10 , X 01, etc. This defines a binary tree structure in which each node has two children. Each series of superscripts defines a unique (L-tuple) node. Our procedure for generating subproblems has three important properties. First, every L-tuple generated from X is either R1 , from X 1 , or from X 0 . Second, no L-tuple generated from X 1 or X 0 has a score greater than score(R1 ). Third, the top-scoring L-tuple from X 0 has the same score as R1 . Given this formulation, a recursive solution to the Top-K problem is theoretically possible. In the base case, the input list X has L vectors of length one, and there is only one possible L-tuple to report (namely, R1 ). Given arbitrary length vectors in X, we can divide the problem as above and merge the resultant top-K lists with R1 , discarding all but the top K elements of the merged list. This method, however, is impractical since each of the lists returned by the recursive calls could contain as many as K elements; hence, violating the requirement that space complexity must not be 1

If all vectors in X contain exactly one element, no subproblems can be generated. In this case, the Top-K problem is trivial.

270

T. Eliassi-Rad and K. Henderson

O(K). But, a search-based approach allows us to generate the top K tuples one at a time. If we treat each Top-K instance as a node, and each subproblem generated from an instance as a child of that node, then we can treat the Top-K problem as search in a binary tree. The cost of traversing an edge is equal to the loss in score incurred by removing elements from X; thus the X 0 edge always has cost 0 and the X 1 edge has cost equal to bestDif f (X) =def min(Di1 |i = 1 · · · L). In this context, A* search [3] clearly generates subproblems in order of their Ri scores. For the heuristic function h(n), we use bestDif f (X), which can be readily computed. Note that bestDif f is monotone (and thus admissible) by the following argument:2 If p is a 1-child of n (the X 1 subproblem), then h(n) = cost(n, p) and h(p) ≥ 0. Otherwise, cost(n, p) = 0 and h(n) ≤ h(p) by the definition of bestDif f . Unfortunately, A* search requires storage of the OP EN list in memory, and the size of OP EN increases with every node expansion. This violates our memory requirements (of less than O(K) storage), so we employ SMA* search [5] which stores a maximum of M nodes in memory at any given time during the execution. SMA* expands nodes in the same order as A* until it runs out of memory. At that point, the least promising node is deleted and its f -value is backed up in its parent.3 SMA* is guaranteed to generate the same nodes as A* and in the same order. However, it may generate some intermediate nodes multiple times as it “forgets” and “remembers” portions of the tree. In certain cases, especially when the parameter M is small compared to the size of the search fringe, SMA* can experience “thrashing.” This thrashing results from large parts of the search tree being generated and forgotten with very few new nodes being discovered. Our serial Top-K algorithm is essentially the same as SMA* [5] except for the use of heaps in selecting which of the Xi vectors to decrement at each node.

3 A Parallel Solution to the Top-K Ranking Problem The aforementioned serial algorithm is very sensitive to the choice of M , the maximum amount of memory that can be allocated (see the Experiments Section). Here we present a parallel SMA* algorithm that offers dramatic improvement in runtime. For a machine with P processing nodes, we use A* search to generate the P –1 best candidate subproblems. This covers the entire search tree. Because each subproblem is independent of the others, each processing node can perform SMA* search on its subproblem and incrementally report results to the master processing node, where the incoming tuple lists are merged and results are reported. For small P , this algorithm works as expected and produces super-linear speedup. However as P increases, the performance boost can decrease quickly. This occurs because the runtime for this parallel algorithm is dominated by the single processing node with the longest runtime. If the initial allocation of subproblems to processing nodes is imbalanced, additional processing nodes may not improve performance at all. To ameliorate this problem, we adopt a load-balancing heuristic. 2

3

Recall that a heuristic function h(n) is monotone if h(n) ≤ cost(n, p) + h(p) for all nodes n and all successors p of n. The function f represents the total cost function, which equals the sum of the cost encountered so far and the estimated cost.

Ranking Information in Networks

271

We run parallel SMA* on the input data with K   K as the threshold parameter. We then use the relative load from each subproblem as an estimate of the total work that will have to be done to solve that subproblem. We use these estimates to redistribute the initial nodes and repeat until there are no changes in the initial allocation of nodes. This distribution is then used to generate the top K L-tuples as described above. In our experiments, this heuristic correctly balanced the loads on the processing nodes. Moreover, the initial overhead to calculate the estimates was a negligible fraction of the overall runtime.

4 Experiments 4.1 Synthetic Data: Results and Discussion To determine the runtime requirements for our Top-K algorithm, we generated a synthetic data set with L = 100 lists. Each list has between one and ten real values distributed uniformly in [0, 1]. Figure 1 shows runtime results for M = 106 nodes in memory. Note that as K increases, time complexity in K becomes near-linear. However, at approximately 20,000 tuples per second it is still too slow to be practical for large values of K. Figures 2(a), 2(b), and 2(c) show the performance metrics for a strong scaling experiment with the parallel algorithm: runtime, speedup, efficiency. Runtime is the wall-clock runtime. Speedup is defined as the time taken to execute on one process (T1 ) divided by the time taken to execute on P processes (Tp ). Linear speedup is the ideal case where 2 processes take half the time, 3 processes take a third of the time, and so forth. Efficiency is defined as speedup divided by the number of processors. Ideal efficiency is always 1, and anything above 1 is super-linear. Values less than 1 indicate diminishing returns as more processors are added. For the parallel Top-K experiments, we use the same 100 lists described above but with M = 105 , K = 2·106, and K  = 104 . There is no I/O time in these experiments, as tuples are simply discarded once they are discovered. Note that the runtime for P = 2 is the same as the serial runtime plus overhead because in this case one processing node is the master and simply “merges” the results from the compute node. We see super-linear speedup for values of P up to 17 processes. However, the runtime for P = 17 is nearly identical to the runtime for P = 9. This is because in the P = 17 experiment, one of the large subproblems is not correctly split and redistributed. Figure 2(d) reveals the cause. When P > 9, the subproblems are not being correctly distributed among the nodes. Figure 3(d) shows the same results for the load-balanced version of the algorithm. In this case, it is clear that the nodes are being correctly distributed. Figures 3(a), 3(b), and 3(c) show the parallel efficiency results for the load-balanced case. Our load-balancing heuristic drastically improves performance as the number of processors increases. These results include the time required to calculate the loadbalance estimates, which explains the slightly longer runtimes at P ≤ 9. 4.2 Real Data: Results and Discussion We tested our Top-K algorithm on four real-world data sets – namely, patent citations, DBLP bibliographic data, IP traffic, and UN voting. For brevity, we only discuss two

T. Eliassi-Rad and K. Henderson

60000

700

50000

600 500

T u p les/Seco n d

40000

400 30000 300 20000

200

10000

T im e (in Seco n d s)

272

100

0

0

0.E+00 1.E+06 2.E+06 3.E+06 4.E+06 5.E+06 6.E+06 7.E+06 8.E+06 9.E+06 1.E+07 K-th L-tuple tuples/sec

total time (secs)

Fig. 1. Serial top-K: Runtime as a function of K

70

10000

50 S p eed u p (T 1 /T P )

R u n tim e (S e c o n d s in L o g S c a le )

60 1000

100

40 30 20

10

10 1 1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

0

18

0

1

2

3

4

5

6

Number of Processors Observed Runtime

8

9

10

11

12

13

14

15

Linear Runtime

Observed Speedup

7

Process Loads (Sorted)

6 5 4 3 2 1 0 2

3

4

5

6

7

8

9

10

11

12

13

Number of Processors Observed Efficiency

Linear Efficiency

(c) Efficiency

18

Linear Speedup

100%

2000

90%

1800

80%

1600

70%

1400

60%

1200

50%

1000

40%

800

30%

600

20%

400

10%

200

0%

1

17

(b) Speedup

8

0

16

Number of Processors

(a) Runtime scaling

E ffic ie n c y (T 1 /(T P *P ))

7

14

15

16

17

18

0 2

3

5

9

17

Number of Processors Runtime

(d) Load balance

Fig. 2. Unbalanced parallel top-K: K = 2 · 106 , L = 100 lists, M = 105 nodes

Time (Seconds)

0

Ranking Information in Networks

273

140

10000

100 S p eed u p (T 1 /T P )

100

80 60 40

10

20 1

0 0

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

0

18

1

2

3

4

5

6

7

Observed Runtime (Balanced)

Observed Runtime (Unbalanced)

Linear Runtime

Observed Speedup (Balanced)

(a) Runtime scaling

9

10

11

12

13

14

15

16

17

Observed Speedup (Unbalanced)

Linear Speedup

9

100%

2000

8

90%

1800

7

80%

1600

70%

1400

60%

1200

50%

1000

40%

800

30%

600

2

20%

400

1

10%

200

0

0%

6 5 4 3

0

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

Observed Efficiency (Unbalanced)

(c) Efficiency

18

0 2

3

5

9

17

Number of Processors

Number of Processors Observed Efficiency (Balanced)

18

(b) Speedup

Proc e s s Loa ds (Sorte d)

E ffic ie n c y (T 1 /(T P *P ))

8

Number of Processors

Number of Processors

T im e (Seco n d s)

R u n tim e (S e c o n d s in L o g S c a le )

120 1000

Linear Efficiency

Runtime

(d) Load balance

Fig. 3. Balanced parallel top-K: K = 2 · 106 , L = 100 lists, M = 105 nodes

of them here. First, our patents data set is a collection of 23,990 patents from the U.S. Patent Database (subcategory 33: biomedical drugs and medicine). The goal here is to discover sets of patents which, if removed, would have the greatest effect on the patent citation network. In particular, we are interested in patents which are frequently members of such sets; and we want to find which centrality measures correspond to these frequent nodes in the citation network. We generated a citation network with 23,990 nodes and 67,484 links; then calculated four centrality metrics on each node: degree, betweenness centrality, random walk with restart score (RWR), and PageRank. In the Top-K framework, there are 4 vectors: one for each centrality measure. The indices into the vectors are patents. The entry i in vector j is the (normalized) centrality score j for patent i. Table 1 lists the patents with the highest frequency in the top-100K tuples. As expected, the patents associated with DNA sequencing have the highest frequencies. Table 2 presents the centrality scores for the patents listed in Table 1. As it can be seen, the centrality scores alone could not have found these highly impacting patents (since no patent dominates all four centrality scores). Our second data set is composed of 4943 proposed UN resolutions, votes from 85 countries per resolution (yes, no, abstain, etc), and 1574 categories. Each proposed resolutions is assigned to one or more categories (with an average of 1.6 categories per proposal). We select the most frequent 20 categories. Then, we score each country in each category by how often it agrees (or disagrees) with the USA vote in that category:

274

T. Eliassi-Rad and K. Henderson Table 1. Patents with the highest frequency in the top-100K tuples

Patent Frequency Patent Title Number (in Top-100K Tuples) 4683195 0.799022 Process for amplifying, detecting, and/or-cloning nucleic acid sequences 4683202 0.713693 Process for amplifying nucleic acid sequences 4168146 0.202558 Immunoassay with test strip having antibodies bound... 4066512 0.175828 Biologically active membrane material 5939252 0.133269 Detachable-element assay device 5874216 0.133239 Indirect label assay device for detecting small molecules ... 5939307 0.103829 Strains of Escherichia coli, methods of preparing the same and use ... 4134792 0.084579 Specific binding assay with an enzyme modulator as a labeling substance 4237224 0.074189 Process for producing biologically functional molecular chimeras 4358535 0.02674 Specific DNA probes in diagnostic microbiology

Table 2. Centrality scores for the 10 highest-frequency patents in the top-100K tuples Patent # 4683195 4683202 4168146 4066512 5939252 5874216 5939307 4134792 4237224 4358535

Degree Betweenness RWR PageRank 1.000 0.433 0.030 1.000 0.974 0.234 0.000 0.975 0.099 1.000 0.132 0.088 0.036 0.845 0.185 0.026 0.221 0.234 1.000 0.000 0.122 0.234 1.000 0.000 0.071 0.234 0.965 0.000 0.126 0.772 0.040 0.118 0.341 0.760 0.016 0.332 0.460 0.670 0.064 0.446

1 m

points per agreement, where m is the number of agreements per proposed resolution. The input to our top-K algorithm then is a matrix F with 85 rows (one per country) and 20 columns (one for each selected category). The entries in this matrix are the weighted agreement scores. We set K (in the top-K) to 100, 000. We repeat the experiment with scores for disagreements. Figures 4(a) and 4(b) depict the ranking results.

5 Related Work Previous studies of the Top-K problem focus on cases where L is small (2 or 3) and each list is very long. Fredman ([2]) studies the problem of sorting X + Y = {x + y|x ∈ X, y ∈ Y } in better than O(n2 · log(n)) time (where |X| = |Y | = n). Frederickson and Johnson ([1]) examine the slightly easier problem of discovering the K th element of X + Y  in O(n · log(n)) time. They also consider the problem of finding the K th m element of i=1 Xi , proving that a polynomial algorithm in n for arbitrary m ≤ n and K would imply that P = N P . Our approach uses fixed-memory heuristic search to enumerate L-tuples. It has O(n · log(n)) runtime and at most O(K) storage.

Ranking Information in Networks +


-


*%3


,)5


*%2






,)4
,)3
,)2
,)1
,)0


*%1
*%0
*%/
*%.
*%-
*%,


,)/


*%+


,).


*


,)-


 






#%#



 

#

'



$





##&



 


$


#&

!

#

$&

# 

#



,



! 



!
& "
'


 !


#
&!'








! 





!


!#



& '


!






!


#


#








$"


 







275

(a) Top countries agreeing with the USA in the (b) Top countries disagreeing with the USA in top-100K tuples the top-100K tuples Fig. 4. UN resolutions: Voting records on the 20 most frequent categories

6 Conclusions We demonstrated that ranking sets of nodes in a network can be expressed in terms of the Top-K problem. Moreover, we showed that this problem can be solved efficiently using fixed memory by converting the problem into a binary search tree and applying SMA* search. We also parallelized SMA* for the Top-K problem in order to achieve super-linear speedup in a distributed-memory environment. Lastly we validated our approach through experiments on both synthetic and real data sets.

Acknowledgments This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under contract No. DE-AC52-07NA27344.

References 1. Frederickson, G.N., Johnson, D.B.: Generalized selection and ranking. In: Proc. of the 12th STOC, pp. 420–428 (1980) 2. Fredman, M.L.: How good is the information theory bound in sorting? Theoretic. Comput. Sci. 1, 355–361 (1976) 3. Hart, P.E., Nilsson, N.J., Raphael, B.: Correction to a formal basis for the heuristic determination of minimum cost paths. SIGART Bull. 37, 28–29 (1972) 4. Russell, S.: Efficient memory-bounded search methods. In: ECAI, pp. 1–5 (1992) 5. Russell, S.J.: Efficient memory-bounded search methods. In: Proc of the 10th ECAI, pp. 1–5 (1992)

Information Provenance in Social Media Geoffrey Barbier and Huan Liu Arizona State University, Data Mining and Machine Learning Laboratory [email protected], [email protected] http://dmml.asu.edu/

Abstract. Information appearing in social media provides a challenge for determining the provenance of the information. However, the same characteristics that make the social media environment challenging provide unique and untapped opportunities for solving the information provenance problem for social media. Current approaches for tracking provenance information do not scale for social media and consequently there is a gap in provenance methodologies and technologies providing exciting research opportunities for computer scientists and sociologists. This paper introduces a theoretical approach aimed guiding future efforts to realize a provenance capability for social media that is not available today. The guiding vision is the use of social media information itself to realize a useful amount provenance data for information in social media. Keywords: provenance, information provenance, data provenance, social media, data mining, provenance path.

1

Introduction and Motivation

Collective behavior can be influenced by small pieces of information published in a social media setting such as a social networking site, a blog, micro-blog, or even a wiki [2,6]. Considering information as defined by Princeton’s WordNet1 as “a collection of facts from which conclusions may be drawn” and provenance as “the history of ownership of a valued object2 ”, information provenance in social media can be considered as the origins, custody, and ownership of a piece of information published in a social media setting. Often the provenance of information published via social media is only known after a group has been influenced and motivated in a particular manner. A lack of accurate, reliable, history or metadata about a social media information source can present problems. In March 2010, John Roberts, a United States Supreme Court Justice, was reportedly set to retire because of health issues. As it turned out, Justice Roberts had no plans to retire and a rumor that grew from a professor’s teaching point, meant only for a classroom, made national headlines [2]. When Twitter was used by protestors in Iran during 2009, the source of 1 2

http://wordnetweb.princeton.edu/perl/webwn http://www.merriam-webster.com/dictionary/provenance

J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, pp. 276–283, 2011. c Springer-Verlag Berlin Heidelberg 2011 

Information Provenance in Social Media

277

some messages could not be verified and were deemed to be of no value or even antagonistic [6]. These problems might have been avoided with provenance information related to the subject, the source, or perhaps even the ideologies in play. Some mechanisms have been designed to record provenance information for databases, the semantic web, workflows, and distributed processing [10]. However, provenance information is not routinely tracked today for social media. Although some thought has been given about the need [7] and some potential approaches [5,11], a practical approach and responsive mechanism has not be identified or implemented for today’s online social media environment. In some instances, partial provenance information may suffice to inform groups in such a manner resulting in sound behaviors. Additionally, an approach for information provenance in social media needs to address the rapidly changing social media environment and must quickly respond to queries about the provenance of a piece of information published in social media. The social media environment provides unique challenges for tracking and determining provenance. First, the social environment is dynamic. With more than half a billion3 Facebook4 users, new social media content is generated every day. Facebook is only one social media outlet. Consider there are over 40 million microblog messages posted everyday to the popular site Twitter5 . Second, social media is decentralized in the sense that information can be published by almost anyone choosing one or more social media platforms and then relayed across disparate platforms to a multitude of recipients. Third, the environment provides multiple modes of communication such as profile updates, blog posts, microblogs, instant messages, and videos. Given this extremely challenging environment, new approaches for information provenance are needed to track where a piece of information came from in order to determine whether or not the information can be used as a basis for a decision. Determining the provenance of a piece of information is challenging because provenance data is not explicitly maintained by most social media applications. A piece of information often moves from one social media platform to another. As a result, provenance data for a piece of information in the social media environment needs to be derived. It is proposed here that social media data itself can be used to derive the provenance or most likely provenance data about a piece of information. This departs from traditional approaches of information and data provenance which rely on a central store of provenance information. The remainder of this paper highlights related work noting at this point in time contemporary approaches to provenance rely on a central store of provenance meta-data or an established, widely adopted, semantic framework. Following the highlights of related work, the concept of a provenance path and associated problem space applicable to today’s social media environment are outlined. An example to illustrate the provenance path concept is included and future research opportunities are presented. 3 4 5

http://www.facebook.com/press/info.php?statistics www.facebook.com http://blog.twitter.com/2010/02/measuring-tweets.html

278

2

G. Barbier and H. Liu

Related Work

Provenance data is valuable in a variety of circumstances including database validation and tracing electronic workflows such as science simulations and information products produced by distributed services. Moreau published a comprehensive survey [10] that captures provenance literature related to the web. Although the survey provides over 450 references with an emphasis on data provenance, the survey does not identify a significant body of literature relating to provenance and social media, social computing, or online social networks. The database and workflow communities depend on a provenance store implemented in one form or another. The proposed store is maintained centrally by one application or in a distributed scheme, each application involved in producing a piece of data maintains its own provenance store. An open provenance model (OPM) was developed to facilitate a “data exchange format” for provenance information [10]. The OPM defines a provenance graph as a directed graph representing “past computations” and the “Open Provenance Vision” requires individual systems to collect their provenance data [10]. Hasan et al. [8] define a “provenance chain.” Deolalikar and Laffitte [3] investigate data mining as a basis for determining provenance given a set of documents. The World Wide Web Consortium has a provenance incubator group6 to develop requirements for the semantic web. Golbeck [5] connects provenance with trust leveraging Resource Description Framework (RDF) supported social networks. Fox and Huang define what they call Knowledge Provenance (KP) [4,9]. Their KP construct accounts for varying levels of certainty about information found in an enterprise (i.e., the web). In today’s social media environment, applications do not collect or maintain provenance data directly. Thus, it would be practical to determine provenance dynamically without depending on a central store. Information provenance in social media must be analyzed, inferred, or mined to be realized.

3

Provenance Paths

Provenance can be characterized as a directed graph [3,5,10,12]. Within the graph, a provenance path can be assembled for each statement produced from the social media environment. We propose next, formal definitions for provenance paths as applied to social media. The social media environment network will be represented by a directed graph G = (V, E), v ∈ V and e ∈ E. V is the set of nodes representing social media users publishing information using social media applications. E is the set of edges in G representing explicit transmission of social media communication between two nodes in V . An explicit transmission occurs when distinct information is communicated from one node to another or when one node directly accesses information available at another node. Publishing information alone is not 6

http://www.w3.org/2005/Incubator/prov/wiki/Main_Page

Information Provenance in Social Media

279

considered an explicit transmission and does not create an edge in E. Given the directed graph G = (V, E). We have the following definitions. Definition: T is the set of recipient nodes in G : T ⊆ V . Definition: A is the set of accepted 7 nodes in G : A ⊆ V and (T ⊂ A). Definition: D is the set of discarded 8 nodes in G : D ⊂ V, (D ∩ A) = ∅, and (D ∩ T ) = ∅. Definition: (A ∪ D) are identified nodes. Definition: M is the set of undecided nodes in G : M = V − (A ∪ D). Definition: a provenance path, p, is a path in G : p = (v1 , v2 , . . . , vn ) : v1 = vn , v1 ∈ V , and vn ∈ T . Definition: P is the set of all provenance paths in G : ∀pi ∈ P, i = 1 . . . m : m = | P | and p1 = p2 = p3 = . . . pm . Definition: Accepted provenance path, p : for all nodes, vk , in path p, vk ∈ A. Definition: Heterogeneous provenance path, p: for all nodes, vj , in path p, vj ∈ A, vj ∈ D, or vj ∈ M . A provenance path is a set of nodes and edges comprising a path which an element of social media information is communicated from a node in graph to a recipient or recipients. Nodes in the set T (an individual or group) are the final recipients of information along a provenance path, hereafter called recipient. The recipient makes one or more decisions based on the information received via Fig. 1. Sets and abstract paths a provenance path. Each provenance path is unique. There may be more than one provenance path for information provided to a recipient. Figure 1 illustrates the most common relationship between the subsets of V . The arrows illustrate some characteristics of possible provenance paths including accepted and heterogeneous provenance paths.

4

Provenance Path Problem Space

The social media environment provides a very challenging problem for determining information provenance. This environment, like the web, provides a theoretically bounded but practically unbounded problem space because of the large 7 8

The criterion for accepting nodes is uniquely determined by T . The criterion for discarding nodes is uniquely determined by T .

280

G. Barbier and H. Liu

number of users in the social media environment. Although there are a finite number of websites as part of the world-wide-web, determining the actual number of web sites is extremely challenging [1]. Similarly, there are a finite number of social media users and social media information. However, the precise actual number of users is practically intractable. This unbounded social media environment presents an unbounded problem space for provenance paths in practice although in some cases the environment will be bounded such as when considering a single social networking site or small subset of social media sites. A provenance path can begin at an identified node (v1 ∈ A or v1 ∈ D) or from a node that is undecided (v1 ∈ M ). The social media environment presents cases where a provenance path exists but all of the nodes and edges in the path are not known or only partially known to the recipient, defined as an incomplete provenance path. In the case of an incomplete provenance path, the complete provenance path exists in the social media environment but the complete path is not discernable to the recipient. Given an incomplete provenance path, the primary goal of solving the provenance path problem is to make all of the unknown nodes and edges known to the recipient. When all of the nodes and edges are known by the recipient, the provenance path is defined as a complete provenance path. When the social media environment is unbounded, it may not be possible to make a complete provenance path known to the recipient. The recipient will need to employ strategies to decide whether or not the revealed portions of the incomplete provenance path provides useful information. The problem space can be considered and approached from different perspectives depending on whether or not the path is complete or incomplete. 4.1

Complete Provenance Paths

Assessing provenance will be easiest when the recipient can access a complete provenance path with all of the node and edge relationships known to the recipient. Identified nodes are determined by a recipient as accepted or discarded. The criterion for accepting nodes can be based on one characteristic or a combination of characteristics attributed to nodes in the social media environment. However, it is intuitive that the most common problem encountered with information provenance in social media is dealing with incomplete provenance paths. 4.2

Incomplete Provenance Paths

If the actual path is not completely known to the recipient, it could be difficult to determine whether or not a discarded node contributed to or altered information presented to the recipient. Social media data could be leveraged to estimate likely paths. The nodes and links that are known to the recipient or consequently discovered can be exploited to provide a warning or calculate confidence values using probabilistic mechanisms to determine how the information might be considered. Approaches need to be designed to infer provenance data when the path is incomplete. Decision strategies need to be developed to help the recipient authenticate information provided through social media or determine whether or

Information Provenance in Social Media

281

not the information itself can be corroborated via a separate provenance path including accepted social media nodes. In some cases it may enough to know whether or not the idea being presented is adversarial, complementary, or unique toward the recipient individual or group. Understanding the nuances of a publication, position, or opinion, could lend itself to a level of confidence acceptable to a recipient by using only the portion of the provenance path that is available for analysis. 4.3

Multiple Provenance Paths

Multiple provenance paths present both prospects and challenges. When multiple provenance paths are complementary, the paths present consistent information to the recipient. Complementary provenance paths can serve as an authentication mechanism for the information presented to the recipient. The most challenging decisions an individual or group may need to make is when the provenance paths are incomplete and multiple provenance paths provide conflicting information. When multiple provenance paths are conflicting by presenting inconsistent or contradictory information to the recipient, the provenance paths must be reconciled. In cases where provenance paths provide conflicting information, a probabilistic approach can be applied to determine which provenance path should be accepted, if any.

5

Provenance Path Case Study

Consider the case of the Justice Roberts rumor based on a simple investigation [2]. Reference Figure 2a, a Georgetown Law School professor (node v1 ) shared fictitious information in his class along edges e1 , e2 , and e3 . A student in the class, node v3 , sends a message to a blog site node v5 along edge e4 and the group at the blog site publishes a story based on false information. Similar provenance paths reach other blog sites and false information about a Justice in the United States Supreme Court becomes a well-circulated rumor. The information communicated along e4 may or may not be accurate. Given the provenance path shown in Figure 2a, node v5 should determine whether or not it should accept the information about Justice Roberts. If the recipient node v5 analyzes the provenance path and determines that it considers each node along the provenance path as accepted, v5 could accept the information received via the explicit communication along e2 and e4 . However, if v1 or v3 are discarded nodes, the recipient will need to consider what must be done in order to authenticate the information. In Figure 2b, an additional node, JR, is added to represent Justice Roberts. If node JR was included the provenance path, the information might be considered reliable. However, given that the node JR is not included in the path (as far as the recipient can initially discern) questions should be raised about the validity of the information. In this case, direct or indirect connections using networking

282

G. Barbier and H. Liu

(a)

(b) Fig. 2. A case study

information and available social media could be examined to glean additional information. As examples, comparing the “distance” from node v1 and node JR to a common reference point in social media or analyzing the individuals’ (v1 , v3 , and JR) group memberships and associated group traits.

6

Future Work

In this paper we defined the problem of information provenance in contemporary social media and presented general ideas for approaching it along with a case study. Our research is ongoing and our future work will be focussed on detailing specific methods and metrics aimed at developing solutions to this problem. In addition to the motivating cases previously discussed, this research has exciting implications for addressing contemporary issues facing users and decision makers such as: identifying the source of an online product review to reveal fake reviews, helping to implement a practical cyber genetics9 , and determining the source when no author is evident. Beyond the work to identify and assess provenance paths there are additional questions related to information provenance in social media such as: – – – –

How should provenance attributes and characteristics be defined? How would provenance paths be valued from different recipients? Can provenance paths be identified and leveraged to help influence a group? In addition to trust, what other connections can be made between provenance and elements of social media? – What are the implications for privacy? Acknowledgments. We appreciate the members of the Arizona State University, Data Mining and Machine Learning laboratory for their motivating influence and thought-inspiring comments and questions with reference to this topic. 9

http://www.darpa.mil/sto/programs/cg/index.html

Information Provenance in Social Media

283

References 1. Benoit, D., Slauenwhite, D., Trudel, A.: A web census is possible. In: International Symposium on Applications and the Internet (January 2006) 2. Block, M.: Tracing rumor of John Roberts’ retirement (March 2010), http://www.npr.org/templates/story/story.php?storyId=124371570 3. Deolalikar, V., Laffitte, H.: Provenance as data mining: combining file system metadata with content analysis. In: TAPP 2009: First Workshop on Theory and Practice of Provenance, pp. 1–10. USENIX Association, Berkeley (2009) 4. Fox, M.S., Huang, J.: Knowledge provenance in enterprise information. International Journal of Production Research 43(20), 4471–4492 (2005) 5. Golbeck, J.: Combining provenance with trust in social networks for semantic web content filtering. In: Moreau, L., Foster, I. (eds.) IPAW 2006. LNCS, vol. 4145, pp. 101–108. Springer, Heidelberg (2006) 6. Grossman, L.: Iran protests: Twitter, the medium of the movement. Time June 17 (2009), http://www.time.com/time/world/article/0,8599,1905125,00.html 7. Harth, A., Polleres, A., Decker, S.: Towards a social provenance model for the web. In: 2007 Workshop on Principles of Provenance (PrOPr), Edinburgh, Scotland (November 2007) 8. Hasan, R., Sion, R., Winslett, M.: Preventing history forgery with secure provenance. Trans. Storage 5(4), 1–43 (2009) 9. Huang, J., Fox, M.S.: Uncertainty in knowledge provenance. In: Bussler, C.J., Davies, J., Fensel, D., Studer, R. (eds.) ESWS 2004. LNCS, vol. 3053, pp. 372–387. Springer, Heidelberg (2004) 10. Moreau, L.: The foundations for provenance on the web. Submitted to Foundations and Trends in Web Science (2009) 11. Simmhan, Y., Gomadam, K.: Social web-scale provenance in the cloud. In: McGuinness, D.L., Michaelis, J.R., Moreau, L. (eds.) IPAW 2010. LNCS, vol. 6378, pp. 298–300. Springer, Heidelberg (2010) 12. Simmhan, Y.L., Plale, B., Gannon, D.: A survey of data provenance techniques. Tech. Rep. IUB-CS-TR618, Computer Science Department, Indiana University, Bloomington, IN 47405 (2005)

A Cultural Belief Network Simulator Winston R. Sieck, Benjamin G. Simpkins, and Louise J. Rasmussen Applied Research Associates, Inc. 1750 Commerce Center Blvd, N., Fairborn, OH, 45324 {wsieck,bsimpkins,lrasmussen}@ara.com

Abstract. This paper describes a tool for anticipating cultural behavior called, “CulBN.” The tool uses Bayesian belief networks to represent distributions of causal knowledge spread across members of cultural groups. CulBN simulations allow one to anticipate how changes in beliefs affect decisions and actions at the cultural level of analysis. An example case from a study of Afghan decision making is used to illustrate the functionality of the tool. Keywords: cultural network analysis, CNA, cultural model, belief network, Bayesian modeling.

1 Introduction The purpose of this paper is to describe a prototype tool for simulating cultural behavior; the “Cultural Belief Network Simulation Tool” (CulBN). We use an example case from a study of Afghan decision making to illustrate key features of the tool and associated process for building cultural models. The tool’s functionality derives from a principled, cognitive approach to culture. We first define and discuss this view of culture, prior to discussing specific details of CulBN. The cognitive approach to culture begins by defining culture in terms of the widely shared concepts, causal beliefs, and values that comprise a shared symbolic meaning system [1, 2]. Shared meaning systems imply that members of a culture tend to perceive, interpret, and respond to events in similar ways. The cognitive approach to culture is often combined with an epidemiological metaphor from which researchers seek to explain the prevalence and spread of ideas in populations [3]. Working from this perspective, we previously developed cultural network analysis (CNA), a method for describing ideas that are shared by members of cultural groups, and relevant to decisions within a defined situation [4]. CNA discriminates between three kinds of ideas: concepts, values, and beliefs about causal relations. The cultural models resulting from CNA use belief network diagrams to show how the set of relevant ideas relate to one another. The CNA approach also includes the full set of techniques needed to build cultural model diagrams. This consists of specific methods to elicit the three kinds of ideas from native people in interviews or survey instruments, extract the ideas from interview transcripts or other texts, analyze how common the ideas are between and within cultural groups, align and assemble the common ideas into complete maps. CNA offers a unique set of theoretical constraints which distinguishes the cultural models it produces from other ways of modeling culture [5]. These aspects include an emphasis on ensuring the relevance of cultural J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, pp. 284–291, 2011. © Springer-Verlag Berlin Heidelberg 2011

A Cultural Belief Network Simulator

285

models to key decisions to provide a more direct link to actual behavior, focus on the cultural insider or emic perspective, modeling interrelated networks of ideas rather than treating ideas as independent entities, and by seeking to directly estimate the actual prevalence of ideas in the network rather than relying on more nebulous notions of sharedness. This overall conception and approach to cultural analysis ideally suits simulations using Bayesian normative modeling, though with a slight twist. In typical applications of Bayesian belief networks (BNs), one elicits expert structural and probability inputs to get best possible representation of a physical system [6]. The aim in such cases is to accurately predict key physical outcomes based on expert understanding of influences within that physical system. That is, the ultimate concern is for physical reality. In contrast, the objective here is to represent human cultural knowledge as the system of interest. That is, the aim is to accurately anticipate the perceptions and decisions of groups within a specified context based on relevant culturally shared conceptions, causal reasoning, and appraised values. The key principles for application of BNs to cultural modeling are: • • • •

BN focuses on population, rather than individual psychological level of analysis Nodes represent concepts held in common by some percentage of the population Edges represent causal beliefs, also distributed across members of the population Probabilities denote prevalence of ideas in the population, not strength of belief

In principle, concept nodes also contain associated value information. However, values are left implicit in the current instantiation. Next, we describe the CNA approach to building and simulating cultural models as a two step process. First, a data step is necessary to translate from raw qualitative data to structured inputs to be used in the second, modeling step. The modeling step incorporates the inputs to build executable models in CulBN. We walk through the two steps in the context of a concrete example of Afghan decision making.

2 Data Step CulBN simulations, as a component of CNA, are grounded in the kinds of data people from a wide variety of cultures, including non-western and non-literate cultures, have been shown capable of providing in natural communications, interviews and field experiments. The first step in a CNA collection is to determine the target population and critical decision or behavior of interest. These elements are often captured in brief scenarios that provide some additional context, such as the following: A unit of American soldiers in Afghanistan’s Uruzgan Valley was engulfed in a ferocious fire fight with the Taliban. Only after six hours, and supporting airstrikes, could they extricate themselves from the valley. The Americans were surprised in the battle by the fact that many local farmers spontaneously joined in, rushing home to get their weapons. The farmers sided with the Taliban, fighting against the Americans. Asked later why they’d done so, the villagers insisted that they didn’t support the Taliban’s ideological agenda, nor were they particularly hostile toward the Americans.

286

W.R. Sieck, B.G. Simpkins, and L.J. Rasmussen

2.1 Interview Process CNA interviews are structured around short scenarios that describe the critical decision of interest for a cultural group, such as the Afghan farmers’ decision to join the fighting in the scenario above. The information basis for the modeling effort described in this paper was a set of CNA interviews conducted with Afghans as part of a larger cultural study [7]. Each participant was interviewed individually using the same CNA interview guide. The interview guide consisted of open-ended questions to elicit participants’ overall explanations of the situation, as well as their perspectives regarding the Afghan protagonists’ intentions, objectives, fundamental values, and the causal links between them. 2.2 Coding and Analysis Two independent coders read through all of the transcripts and identified segments that contained concept-causal belief-value (CBV) chains. Next, the two analysts coded each segment by identifying the antecedent, the consequent, and the direction of the relationship between the antecedent and consequent for each causal belief. For example: “Americans bring security” (concept) “decreases the likelihood that” (causal belief) “Farmers join the Taliban side of the fight” (concept). The coders then collaborated to resolve overlaps and redundancies between their respective lists. Percent agreement for the coding was 81% for the Farmers scenario. Prevalence information was then derived by tallying the number of times a CBV chain was mentioned across all three interviews. We converted the frequencies to proportions by using the count of the most prevalent idea as the upper bound. Finally, we created a frequency table for all ideas. The CBV chains were then integrated into a network diagram of the Afghan Farmers’ decisions leading to their behavior in the scenario.

3 Modeling Step In the modeling step, we construct the cultural belief network in CulBN, and use it as a starting point from which to perform simulations of concept and causal belief change on the decision of interest. The model is initialized with the empirical data that was collected above in the data step. The full cultural belief network for the Afghan Farmers is shown in Figure 1. Recall that these models represent how the cultural group perceives reality. They do not attempt to model physical reality. Hence, we can change various parameters of the model to simulate effects of changes in the distribution of ideas on cultural decisions and behavior. The remainder of this section describes the elements and process of the modeling step in more detail. 3.1 Diagram Semantics in CulBN A subset of the full model is illustrated in Figure 2 to enhance readability. In these diagrams, nodes represent culturally-shared concepts, and the probability on the node

A Cultural Belief Network Simulator

287

Fig. 1. Cultural belief network of Afghan Farmers’ decision making in a conflict scenario as implemented in CulBN

signifies the prevalence of the concept in Afghan thinking in the scenario context. The solid arrows represent perceived positive causal relationships and dotted arrows represent negative ones. Probabilities on the lines denote prevalence of the causal beliefs. As shown in Figure 2, for example, the prevalence of the causal belief that jobs increase wealth for Afghans is 38%. The values the model takes as input are initially determined empirically. These values represent the current state of the cultural model, as derived from the text data. To perform a simulation, the user revises the inputs with hypothetical values. The distinction is visualized in Figure 2. Here, the prevalence of the concept 'US Brings Security' is low, based on the data. However, we may wish to explore the consequences of a case where the concept in much more prevalent in the cultural group (as a result of radio announcements about security, for example). In Figure 2, we see the initial empirical value is represented by the dotted level on the node and the hypothesized values are represented with the solidfill.

288

W.R. Sieck, B.G. Simpkins, and L.J. Rasmussen

After specifying input values, the user runs the simulator to acquire the probability values of the dependent nodes1. An output probability can be interpreted as the prevalence of the concept in the cultural group given the input values specified. For example, in Figure 2 we see that 'Join Taliban Side' is less prevalent with the increased perceptions that the ‘US brings security.’

Fig. 2. Hypothetical and empirical values in CulBN

3.2 Interaction with Cultural Models in CulBN CulBN contains a user interface that allows for creation, visualization, and manipulation of input values in a CNA-type cultural model. Combining this capability with the simulation engine allows the user to easily interact with a cultural model by supplying hypothetical data and visualizing the changes produced in the model. Figure 3 demonstrates this in the Afghan Farmer example. In the top left, we see a model initialized with only empirical data being run (the small dialog is the sample count). To the right, we see the output of the model that 75% of Afghan farmers are expected to join the Taliban in the scenario. In the bottom panels, we explore the effects of hypothesized changes to cultural concepts and causal beliefs related to perceptions of the extent that Americans are building schools locally. First, we hypothesize an increase in the prevalence of perceptions of school-building in the 1

CulBN currently relies on the Java Causal Analysis Tool (JCAT) as its simulation engine 8.Lemmer, J. Uncertainty management with JCAT. 2007 [cited 2010; Available from: www.au.af.mil/bia/events%5Cconf-mar07%5Clemmer_jcat.pdf.. JCAT back-end uses a Monte Carlo Markov Chain (MCMC) sampler to stochastically generate the outcome likelihoods.

A Cultural Belief Network Simulator

289

population (see the bottom left of Figure 3). Note that this does not result in much change in the percentage of farmers who decide to join the Taliban (notice that the solid fill on the 'Join Taliban Side' node is almost the same as the dotted fill). The implication here is that, based on the current model, influencing perceptions that schools are being built is not an effective strategy, in and of itself. However, if we incorporate an additional presumed influence on the causal belief that American school building leads to more jobs, then we will notice an impact on the decision to join the Taliban (see the bottom right of Figure 3). This example shows that subtle distinctions in the specific cultural ideas targeted for influence can yield sizable differences in the predicted effects on the decisions underlying cultural behavior.

Fig. 3. Inputting hypothetical data and running the simulator in CulBN. Top left- running the empirical model, top right- result of empirical model, bottom left-result of a run with hypothetical input to a concept, bottom right- result of second run with hypothetical input added to links.

4 Discussion 4.1 Practical Implications CulBN and its underlying principles are likely to prove useful in a wide array of applications. Here, we describe benefits of the tool for supporting analysts who need to understand the behavior of groups from other cultures. First, CulBN provides a

290

W.R. Sieck, B.G. Simpkins, and L.J. Rasmussen

coherent approach for constraining and making sense of a vast array of concepts, causal beliefs, and values, as they are relevant within particular situations. CulBN, relying on the CNA process, begins with a group’s key decisions, attitudes, or behaviors of interest in a given context. It then provides a systematic approach for identifying and linking the most relevant concepts, causal beliefs, and values from the group’s own perspective to the key decisions. Secondly, CulBN yields a nuanced perspective of the thinking in diverse cultural groups. This is especially helpful to analysts who have little immersive experience within the culture. By providing an explicit map of the group’s perceived associations between concepts, the analyst has a ready reference that describes “what else members of the group would be thinking” in a given situation. Thus, it provides direct support for taking the group’s perspective in that situation. Thirdly, CulBN increases the systematicity and credibility of analyst recommendations, while maintaining flexible analysis processes. CulBN introduces a rigorous process for structuring available information regarding the cognitive content shared among members of a group, allowing analysts to freely explore the cognitive effects of hypothetical influences on the cultural group of interest. Fourthly, CulBN aids in promoting analysts’ cultural expertise across teams and organizations. The cultural model representations are useful knowledge products that can be shared among analysts to provide a concrete basis for discussions, including helping novice analysts get up to speed on specific topics relevant to their area of study. In addition, CulBN could be used to monitor changes in a group’s, organization’s, or wider society’s attitudes and behaviors over time. That is, cultural models provide a basis for representing long-term cultural shifts in attitude and belief. 4.2 Theoretical Implications The current work extends past developments in cultural network analysis by deriving quantitative cultural models from qualitative data, and incorporating them into a BN framework that enables simulated influence on cultural perceptions, decisions, and behavior. There are several advantages of the current approach to achieve cultural simulations. First, cultural agents that do not have the empirical grounding that comes by incorporating quantitative cultural models cannot be guaranteed to behave in culturally authentic ways. The performance of any model is bound by the data used to inform it. Second, a BN fits closely with the representational commitments made within CNA, allowing for the specific interpretations of elements and parameters that fit neatly within the epidemiological view of culture. Also, the quantitative cultural models incorporated into CulBN address the cultural “levels of analysis problem” that exists with many agent-models that aim to represent culture [9]. The problem is that the “cultural” level of analysis describes population characteristics, whereas cognitive agents represent individual people. By building and simulating quantitative cultural network models that estimate population-level characteristics of cultural groups, we can avoid building agents that stereotype the culture of interest. That is, CNA, combined with tools like CulBN, provides a coherent, grounded, and feasible means to model and simulate effects of changes in the distribution of mental representations in a population, as an alternative to stereotyped cultural agents that represent a canonical form of the culture.

A Cultural Belief Network Simulator

291

Acknowledgments This research was sponsored in part by the U.S. Army Research Laboratory and the U.K. Ministry of Defence and was accomplished under Agreement Number W911NF-06-3-0001. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the U.S. Army Research Laboratory, the U.S. Government, the U.K. Ministry of Defence or the U.K. Government. The U.S. and U.K. Governments are authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation hereon. The data used was collected under CTTSO/HSCB contract W91CRB-09-C-0028. Special thanks to John Lemmer, Ed Verenich, and Michael Dygert for the JCAT API and support, and to Shane Mueller and David Oakley for helpful comments.

References 1. Rohner, R.P.: Toward a conception of culture for cross-cultural psychology. Journal of Cross-Cultural Psychology 15(2), 111–138 (1984) 2. D’Andrade, R.G.: The cultural part of cognition. Cognitive Science 5, 179–195 (1981) 3. Sperber, D.: Explaining culture: A naturalistic approach. Blackwell, Malden (1996) 4. Sieck, W.R., Rasmussen, L.J., Smart, P.: Cultural network analysis: A cognitive approach to cultural modeling. In: Verma, D. (ed.) Network Science for Military Coalition Operations: Information Extraction and Interaction, pp. 237–255. IGI Global, Hershey (2010) 5. Sieck, W.R.: Cultural network analysis: Method and application. In: Schmorrow, D., Nicholson, D. (eds.) Advances in Cross-Cultural Decision Making, pp. 260–269. CRC Press / Taylor & Francis, Ltd., Boca Raton (2010) 6. Edwards, W.: Hailfinder: Tools for and experiences with Bayesian normative modeling. American Psychologist 53(4), 416–428 (1998) 7. Sieck, W.R., et al.: Honor and integrity in Afghan decision making. Manuscript in preparation (2010) 8. Lemmer, J.: Uncertainty management with JCAT (2007), http://www.au.af.mil/bia/ events%5Cconf-mar07%5Clemmer_jcat.pdf (cited 2010) 9. Matsumoto, D.: Culture, context, and behavior. Journal of Personality 75(6), 1285–1320 (2007)

Following the Social Media: Aspect Evolution of Online Discussion Xuning Tang and Christopher C. Yang College of Information Science and Technology Drexel University {xt24,chris.yang}@drexel.edu

Abstract. Due to the advance of Internet and Web 2.0 technologies, it is easy to extract thousands of threads about a topic of interest from an online forum but it is nontrivial to capture the blueprint of different aspects (i.e., subtopic, or facet) associated with the topic. To better understand and analyze a forum discussion given topic, it is important to uncover the evolution relationships (temporal dependencies) between different topic aspects (i.e. how the discussion topic is evolving). Traditional Topic Detection and Tracking (TDT) techniques usually organize topics as a flat structure but it does not present the evolution relationships between topic aspects. In addition, the properties of short and sparse messages make the content-based TDT techniques difficult to perform well in identifying evolution relationships. The contributions in this paper are two-folded. We formally define a topic aspect evolution graph modeling framework and propose to utilize social network information, content similarity and temporal proximity to model evolution relationships between topic aspects. The experimental results showed that, by incorporating social network information, our technique significantly outperformed content-based technique in the task of extracting evolution relationships between topic aspects. Keywords: Social Network Analysis, Aspect Evolution Graph, Social Media.

1 Introduction With the advance development of Web 2.0 technology, social networking websites, e.g. Facebook, Twitter, MySpace and other online communities, have become a prevailing type of media on the internet. Users consider these websites not only as networking hubs, but also as ideal sources of information and platforms of exchanging ideas. For instance, thousands of threads are developed in MySpace’s news and politics forum along with the development of daily worldwide news events. However, due to the increasing number of threads and their diversity, it is extremely difficult for a user to follow these threads and fully understand the underlying blueprint. The goal of this paper is to organize threads of a given topic in an easy-tofollow fashion which reflects the evolution relationships between topic aspects, demonstrates the blueprint, and offers richer information. A promising and effective way of organizing threads for capturing the underlying structure and easy navigation is the evolution graph, as shown in Figure 1. Given a topic of interest and a collection of threads about the topic, closely related threads can J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, pp. 292–300, 2011. © Springer-Verlag Berlin Heidelberg 2011

Following the Social Media: Aspect Evolution of Online Discussion

293

Fig. 1. Process of organizing threads from online forum into aspect evolution graph

be grouped together as topic aspects (an aspect is defined as a facet or specific subtopic of a topic here), and pairs of topic aspects with evolution relationship are linked together to form an aspect evolution graph. Organizing threads based on topic aspects and modeling the evolution relationships between aspects are nontrivial tasks. Research in Topic Detection and Tracking (TDT) focused on the automatic techniques of identifying new events and organizing them in a flat or hierarchical structure based on their topics. However, the flat or hierarchical representation of events adopted in TDT cannot illustrate the development process of incidents [1]. To tackle this problem, Nallapati etc. [1] proposed the concept of event threading which captured the dependencies among news events. Yang etc. [2] further proposed the concept of event graph and the techniques of constructing the graph from news corpora. However, these techniques cannot be applied directly to solve the related problem in online forum, because they relied heavily on the content-based analysis techniques. Since most messages in an online forum are short and sparse, traditional TDT techniques cannot capture the topics of these threads effectively based on their content. Furthermore, a thread is very different from a news article because a news article often describes only one single story but a thread can involve multiple stories and opinions. In this paper, our major contribution is to propose a framework of constructing aspect evolution graphs for online forum, which organizes forum threads in a clear and easy-to-follow fashion. We first cluster related threads into topic aspects, and then employ hybrid metrics to detect evolution relationship between topic aspects. Finally, we polished the evolution graph automatically by graph pruning method. To our best knowledge, there is no existing work of tracking aspect evolution and building evolution graph in the context of online forum. We conducted intensive experiments on forum data. Experimental results indicate the effectiveness of our proposed techniques.

2 Related Work Topic Detection and Tracking (TDT) is one line of research closely related to our work, which focuses on capturing topics from a collection of documents. It also

294

X. Tang and C.C. Yang

considers how to cluster documents into different topics following either a flatstructure [3] or a hierarchical structure [4]. However, the main task of TDT does not include evolution relationship modeling as we address in this paper. Topic evolution is another line of research which focuses on how to model changes of topics over time. Some research works study topic evolution by clustering approach[5, 6]. Some other research works study topic evolution by generative models [7-9]. Although these works had their strengths on modeling dynamic topics, they focused more on topics modeling instead of the transitional flow of topic aspects. In addition, these techniques were introduced to study news article or scientific literature, their performance on forum-like data cannot be guaranteed since the forum messages are usually sparse and short. There are a few research works concentrating on modeling event evolution on news articles. Nallapati et al. defined the concept of event threading [1]. They proposed to use a tree structure to capture the dependencies among events in news articles. The similarity between two events in their work was mainly defined by content similarity, location similarity and person name similarity. Following their work, Yang, Shi and Wei [2] introduced the event graph to describe dependencies among events in news article. They considered not only content similarity but also temporal proximity and document distributional proximity to model the relationships between events in an incident. However, none of these works consider modeling evolution relationships in the context of online forum. Forum-like data is usually short, sparse and unformatted so that it is more challenging to deal with comparing to news articles. Moreover, these existing works did not take social network information into account which can be critical in identifying the evolution relationship between topic aspects in an online forum.

3 Problem Formulation In this section, we present the definitions of thread, aspect, topic and aspect evolution graph. Definition 1 (Thread): A thread is a collection of user contributed messages. Its attributes include a title, timestamps, an original post which opens dialogue or makes announcement by the poster, and multiple replies from other forum members. A thread is the smallest atomic unit in our research. We represent a thread with a tuple , , , where is TF-IDF term vector , ,…, | | , and is the original is the social network (defined below) associated with posting time of . Definition 2 (Social Network): A social network associated with a thread is a , , where is a set of users, , , … , | | , participated in graph , and denotes replying relationship between the users of . An edge can be either undirected or directed. In this work, we only consider the directed edge. ,

exists iff

the number of times

replied to replying to

in thread . The weight of this edge equals to in .

Following the Social Media: Aspect Evolution of Online Discussion

295

Definition 3 (Aspect): Given a collection of threads on the topic of interest, these threads can involve in different aspects (facets or subtopics) on the topic. In this paper, each aspect consists of a set of non-overlapping threads, denoted by , , with is the centroid of the TF-IDF term vectors of the threads belonging to

,

|

|

∑

,

is the social network combining all

with by taking union and min | . It’s easy to extend our definition of aspect from non-overlapping threads to overlapping threads. Definition 4 (Topic): A topic consists of a group of related aspects. It represents the common theme of these related aspects. For example, threads about Copenhagen Climate Conference, Glacier Gate and Al Gore’s Nobel Prize are discussing two different aspects of a common topic of global warming. Definition 5 (Aspect Evolution Graph): An aspect evolution graph is a representation of a given topic with several aspects, where its nodes are representing the aspects and its edges are representing the evolution relationships among these aspects. An evolution relationship between two aspects in an evolution graph represents a temporal logical dependencies or relatedness between two aspects. We represent the aspect evolution graph by a tuple , , where , , … , | | is a collection of aspects of a given topic, R is a set of evolution . relationships (directed edges) among The major difference between this work and the previous works such as TDT and event threading is that we focus the modeling effort on evolving aspects in an online forum by making use of social network information rather than events in news articles. Based on the definitions of these concepts, we formalize the major steps of aspect evolution graph modeling (AEGM) as: first of all, given a collection of threads of the same topic, , ,…, , k major aspects, , ,…, , are created where each aspect consists of one or more threads; secondly, given a set of aspects A, we compute the likelihood of evolution relationships between any pairs of aspects in A and then construct an aspect evolution graph; at third, graph pruning method removes those directed edges corresponding to invalid or weak aspect evolution relationships and then generates the final aspect evolution graph (AEG)

4 Modeling Aspect Evolution Graph with Hybrid Model 4.1 Aspect Extraction by Clustering Every topic of discussion consists of multiple aspects. As a result, discussion threads should be clustered into aspects before modeling the evolution relationships among them. In this work, we consider the arriving threads as a finite sequence and infer the model parameters under the closed world assumption. For our online forum scenario, each thread is represented by a TF-IDF term vector. We propose to use the single-pass clustering algorithm [10] based on the TF-IDF term vectors of threads. The singlepass clustering algorithm is utilized for the following reasons: (1) we can apply single-pass clustering on the whole collection of threads without partitioning threads

296

X. Tang and C.C. Yang

by discretizing the time dimension into several time periods; (2) the single-pass clustering does not require prior knowledge, such as total number of clusters like KMeans in which the number of aspects is impossible to be pre-determined for forum threads; (3) the single-pass clustering is more efficient than other clustering approaches which require the computation of the similarity between all pairs of objects, such as hierarchical clustering; (4) it is natural and logical to cluster threads incrementally according to their posting time. 4.2 Modeling Evolution Relationship Given a collection of aspects generated by single-pass clustering algorithm, we propose to utilize text information, temporal information, and social network information to model the evolution relationships between them. 4.2.1 Content-Based Measurement As described in [2], a simple way to estimate the likelihood that an aspect evolution relationship exists from an aspect to is based on the content-based similarity and temporal information. As given in Definition 3 of Section 3, a thread is represented as a TF-IDF term vector and an aspect is represented by the centroid of the term vectors in the aspect. Thus, the content similarity between two aspects is the cosine similarity , . Other than the contentof their centroid term vectors, defined by cos based similarity, the temporal order and temporal distance between two aspects are also useful to determine the evolution relationship. We assume that if one aspect i , it is impossible to have an occurs later than another aspect j evolution relationship from i to j. Moreover, if two aspects are far away from each other along the timeline, the evolution is less likely to happen between them. On the contrary, if two aspects are temporally close, it’s more likely that one will evolve to another. Temporal proximity is used to measure relative temporal distance between two aspects, defined as the follow decaying function: ,

,

/

where is a time decaying factor between 0 and 1, diff(•) denotes the absolute value of time difference between two aspects, T denotes the temporal distance between the start time of the earliest aspect timestamp and the end time of the latest aspect timestamp in the same topic. Based on the notations above, the likelihood of aspect evolution relationship from to is defined as: ,

0 cos

,

,

4.2.2 Content-Based Measurement + Local Social Network Analysis Although content-based similarity is an effective way to capture the evolution relationships between events in news corpora, it cannot guarantee a satisfactory performance on the forum data. Unlike news articles, forum messages are very short and sparse, which make the text similarity measurements less effective. On the other hand, if two aspects are closely related, users who are active in discussing one aspect

Following the Social Media: Aspect Evolution of Online Discussion

297

will tend to be active in discussing another aspect. Thus, two aspects attracting similar groups of users are more likely to be related and evolved from one to another. On the contrary, if two aspects share relatively high content-based similarity but few overlapping between their participants, it is less likely that these two aspects are related. Following this argument, we propose measurements by regularizing contentbased measurement with social network analysis techniques. To regularize the content-based measurement with the intersection of important participants involving in two aspects, we proposed the Local Social Network Analysis technique. It is worth emphasizing that, instead of counting the number of overlapping participants in both aspects only, we assign a significant score to each individual participant of the aspect. Two groups of participants belonging to two aspects can be considered as overlapping if most of the important participants overlap, even though many unimportant participants are different. A formal definition , between two aspects and as: of ,

∑

min

,

∑

max

,

represents the significant score of the participant in aspect , denotes the intersection of participants of two social networks, and similarly denotes the union of participants of two social networks. From the definition above we can learn that the larger the intersection of participants involving in both aspects, the higher the value of Overlap(•,•). Moreover, given a participant involving in both aspects, the smaller the difference between the significant scores of this participant in two social networks, the higher the value of Overlap(•,•). In this work, we explore different ways of measuring significance of an individual participant in a social network, which includes: IN-DEGREE, OUT-DEGREE, INDEGREE PLUS OUT-DEGREE, BETWEENNESS, CLOSENESS and PAGERANK. Based on the notations above, the likelihood of aspect evolution and regularizing by Local Social Network Analysis is relationship between defined as: where

score A , A 0 ε

where

cos W , W

1

ε

Overlap N , N

if opt

opt

tp A , A if opt

opt

are the networks associated with aspects

and 0

1.

4.3 Pruning Aspect Evolution Graph Following the techniques proposed in Section 4.2, an aspect evolution graph with a high density of edges is constructed. A pruning technique is needed to remove those directed edges corresponding to invalid or weak aspect evolution relationships by considering the weights of edges. In this work, we simply set a static threshold for graph pruning. Every relationship with weight lower than the threshold will be removed.

298

X. Tang and C.C. Yang

5 Datasets In this section, we present experiments to evaluate the aspect evolution graph modeling (AEGM) framework. We built our own crawler and retrieved threads in the topics of Haiti earthquake, Nobel Prize and Global Warming by keyword searching. We constructed three datasets for these topics, which include: half month of threads discussing Haiti earthquake (HE dataset), 3 months of threads about President Obama wins Nobel Prize (NP dataset), and 3 years of threads discussing Global Warming (GW dataset). HE dataset contains 17 threads and 1,188 messages. NP dataset consists of 35 threads and 2268 messages. GW dataset contains 36 threads and 5627 messages. After crawling all related threads of these three topics. We solicited human annotators to cluster threads into aspects and then create evolution relationships between these aspects. The aspect evolution graphs created by human annotators are considered as ground truth for comparison in the experiment part.

6 Experiments and Analysis 6.1 Evaluation Measures Taking the aspect evolution graphs , generated by human annotators as the ground truth, we employed Precision, Recall and F-Measure to evaluate the , , which is created automatically by our aspect evolution graph proposed techniques. Defined formally, Precision, Recall and F-Measure can be |/| | , | |/| |. computed as follows: | 6.2 Experiments As introduced in section 4.2.2, we combine Overlap function with text similarity to measure the likelihood of evolution relationship between two aspects. We proposed different possible ways of measuring the significant of individuals in this Overlap function. Table 1. Evolution relationship detection on HAITI dataset (HE), Nobel Prize dataset (NP) and Global Warming dataset (GW) pruned by static thresholding with 0.1

Text TI TO TIO TC TB TP

Best Parameter HE NP =1 =1 =0 =0.8 =0.2 =0.8 =0.5 =0.7 =0.1 =0.8 =0.5 =0.8 =0 =0.8

Precision GW =1 =0.9 =0.7 =0.7 =0.1 =0.8 =0.8

HE 0.45 0.73 0.57 0.56 0.77 0.67 0.59

NP 0.41 0.58 0.57 0.53 0.59 0.5 0.67

Recall GW 0.46 0.27 0.61 0.58 0.55 0.63 0.55

HE 0.87 0.53 0.87 0.93 0.67 0.67 0.87

NP 0.3 0.65 0.7 0.74 0.7 0.35 0.7

GW 0.58 0.63 0.58 0.58 0.58 0.53 0.63

HE 0.59 0.62 0.68 0.7 0.71 0.67 0.7

F Measure NP 0.35 0.61 0.63 0.62 0.64 0.41 0.68

GW 0.51 0.57 0.59 0.58 0.56 0.57 0.59

In the experiment, we attempted to compare the effectiveness of different aspect evolution relationship scoring functions, including: (1) TEXT: Temporal proximity

Following the Social Media: Aspect Evolution of Online Discussion

299

multiplied by aspect content similarity; (2) TEXT+INDEGREE (TI): Temporal proximity multiplied by weighted sum of aspect content similarity and value of Overlap function using INDEGREE as significant score; (3) TEXT+OUTDEGREE (TO): Temporal proximity multiplied by weighted sum of aspect content similarity and value of Overlap function using OUTDEGREE as significant score; (4) TEXT+INDEGREE+OUTDEGREE (TIO): Temporal proximity multiplied by weighted sum of aspect content similarity and value of Overlap function using INDEGEE+OUTDEGREE as significant score; (5) TEXT+CLOSENESS (TC): Temporal proximity multiplied by weighted sum of aspect content similarity and value of Overlap function using CLOSENESS as significant score; (6) TEXT+BETWEENNESS (TB): Temporal proximity multiplied by weighted sum of aspect content similarity and value of Overlap function using BETWEENNESS as significant score; (7) TEXT+PAGERANK (TP): Temporal proximity multiplied by weighted sum of aspect content similarity and value of Overlap function using PAGERANK as significant score. Due to the limited space, we cannot present all of the results in parameter tuning. In Table 1, we summarized the best performance of our proposed techniques with the optimal value of the parameter . For each dataset, we highlighted two techniques with the highest F-Measure. Based on the results, TEXT+CLOSENESS and TEXT-PAGERANK achieved the best performance. TEXT+OUTDEGREE was slightly behind.

7 Conclusion In this paper, we have investigated a new perspective of organizing forum threads and tracking evolving aspects in an online forum. We have presented the frame work of aspect evolution graph modeling for constructing aspect evolution graphs automatically from a collection of threads. By incorporating social network analysis techniques, we are able to address the weakness of content-base techniques. We proposed various approaches of combining content-based and network-based techniques of predicting evolution relationship between two aspects. The experimental results demonstrated a significant improvement of our methods. Future work can include designing more effective approach to combine content-based and network-based techniques to further improve prediction accuracy, or developing more powerful algorithm to cluster threads into aspects automatically.

References 1. Nallapati, R., Feng, A., Peng, F., Allan, J.: Event threading within news topics. In: Proceedings of the Thirteenth ACM International Conference on Information and Knowledge Management, pp. 446–453. ACM, New York (2004) 2. Yang, C., Shi, X., Wei, C.: Discovering event evolution graphs from news corpora. IEEE Transactions on Systems, Man, and Cybernetics, Part A: Systems and Humans 39, 850–863 (2009) 3. Allan, J., Carbonell, J., Doddington, G., Yamron, J., Yang, Y.: Topic detection and tracking pilot study: Final report. In: Proceedings of the DARPA Broadcast News Transcription and Understanding Workshop, vol. 1998, pp. 194–218 (1998)

300

X. Tang and C.C. Yang

4. Allan, J., Feng, A., Bolivar, A.: Flexible intrinsic evaluation of hierarchical clustering for TDT. In: Proceedings of the Twelfth International Conference on Information and Knowledge Management, pp. 263–270. ACM, New York (2003) 5. Morinaga, S., Yamanishi, K.: Tracking dynamics of topic trends using a finite mixture model. In: Proceedings of the Tenth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 811–816. ACM, Seattle (2004) 6. Schult, R., Spiliopoulou, M.: Discovering emerging topics in unlabelled text collections. In: Manolopoulos, Y., Pokorný, J., Sellis, T.K. (eds.) ADBIS 2006. LNCS, vol. 4152, pp. 353–366. Springer, Heidelberg (2006) 7. Mei, Q., Zhai, C.: Discovering evolutionary theme patterns from text: an exploration of temporal text mining. In: Proceedings of the Eleventh ACM SIGKDD International Conference on Knowledge Discovery in Data Mining, pp. 198–207. ACM, New York (2005) 8. Blei, D.M., Lafferty, J.D.: Dynamic topic models. In: Proceedings of the 23rd International Conference on Machine Learning, pp. 113–120. ACM, Pittsburgh (2006) 9. Wang, X., McCallum, A.: Topics over time: a non-markov continuous-time model of topical trends. In: Proceedings of the 12th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 424–433 (2006) 10. Yang, Y., Pierce, T., Carbonell, J.: A study of retrospective and on-line event detection. In: Proceedings of the 21st Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, pp. 28–36. ACM, Melbourne (1998)

Representing Trust in Cognitive Social Simulations Shawn S. Pollock, Jonathan K. Alt, and Christian J. Darken MOVES Institute, Naval Postgraduate School Monterey, CA 93943 {sspolloc,jkalt,cjdarken}@nps.edu

Abstract. Trust plays a critical role in communications, strength of relationships, and information processing at the individual and group level. Cognitive social simulations show promise in providing an experimental platform for the examination of social phenomena such as trust formation. This paper describes the initial attempts at representation of trust in a cognitive social simulation using reinforcement learning algorithms centered around a cooperative Public Commodity game within a dynamic social network. Keywords: trust, cognition, society.

1 Introduction One of the fundamental phenomena governing human interactions is the notion of trust, without which the fabric of society quickly comes unraveled. Trust in other humans and societal institutions facilitate a market economy and a democratic form of government. The human information processing system, adept at receiving and synthesizing large amounts of sensory information from our environment, manages to identify which percepts are salient to our current task or our long term well being by the allocation of selective attention [1]. In this view the level of trust is a quality associated with each percept either directly, one sees an event as a first person observer, or indirectly, the information comes from a reliable source. The latter case, involving the interaction and development of trust between individuals, will be the topic of this paper. The representation of trust within cognitive social simulations is of fundamental importance to the exploration of macro level social phenomena. When humans interact with one another there is a multitude of information, both verbal and nonverbal, that is exchanged between the participants. In social simulations the major issue in modeling this phenomenon is to understand more fully how information flows through a social network and how the society develops and evolves its beliefs over time [2]. Central to the modeling of information flow in social systems is the concept of trust. The main contribution of this paper is a model of trust based on reinforcement learning and demonstrated in the context of the Public Commodity Game [3]. This paper first provides a brief introduction to trust, reinforcement learning, and cognitive social simulation. This is followed by a description of the general trust model as well as the current Python implementation of the model within the context J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, pp. 301–308, 2011. © Springer-Verlag Berlin Heidelberg 2011

302

S.S. Pollock, J.K. Alt, and C.J. Darken

of the Public Commodity game. Computational results of experimentation with model parameters related to the formation of norms and penalties for the violation of norms are provided as well as a summary and discussion of future work.

2 Background This section provides an overview of the notion of trust as used in this setting, reinforcement learning, and cognitive social simulation. 2.1 Trust In order to model trust, it is first necessary to define trust. Trust is not a simple concept (though it is easy to express its general intent) and codifying trust to the extent of incorporation into computer simulation is rather difficult. This work establishes a definition of trust suitable for use in simulation rather than a general definition. In particular the definition of trust used is chosen with an eye toward the long term goal of the research, to model communication and belief revision. Trust is viewed as an analysis performed by an individual agent to prejudice newly obtained information either based on the sender or the topic of discussion, and the history between the sender and the receiver and topic [4]. As an example, if a person who is completely trusted by another shares some information, the receiver is more likely to welcome the communication and take action on it. Additionally, if the sender is moderately trusted by the individual, but the receiver is untrusting of the topic (such as a particular political view), they may disregard the information as non-actionable. Simply making a binary decision of actionable versus non-actionable information is not robust enough to capture the nuances of human behavior requiring the incorporation of some concept of the grey area in between. The model incorporates sender and topic into the state space of a reinforcement learning algorithm to develop a two-pass notion of trust where the receiving agent first determines whether information is actionable then makes a separate decision if he should or should not revise his beliefs based on this new information. 2.2 Reinforcement Learning An appealing approach to represent human like learning and action selection is the idea of reinforcement learning, where agents will seek select actions within their environment based on their experience. Based on the permissiveness of the environment, agents are eligible to receive percepts from the environment that inform them on the state of the environment at a given point in time. The basic elements of reinforcement learning are: a policy that maps states to actions; a reward function that maps a state of the environment to a reward; a value function that maps states to long term value given experience; and an optional model of the environment [5]. The policy provides a set of actions that are available in a given state of the environment; the agents leverage their prior knowledge of the environment, informed by the value function, to determine which action will provide the greatest reward, as defined by the modeler. Agents must strike a balance between exploration, behavior to explore the reward outcomes of state action pairs that have not been tried, and exploitation,

Representing Trust in Cognitive Social Simulations

303

behavior that takes advantage of prior knowledge to maximize short term rewards, in order to avoid converging to local minima [5]. The ability to control this balance makes reinforcement learning an attractive approach for representing human behavior. The reinforcement learning technique used in this work is Q-learning in conjunction with a softmax function (the Boltzmann distribution). Q-learning, Q(s,a)←Q(s,a)+α(r+γmaxa'Q(s',a')-Q(s,a)), falls into a class of model free reinforcement learning methods that have the property that the learned actionvalue function, Q, approximates the optimal action-value function, Q*, requiring only that all state action pairs be updated as visited [5]. For each state action pair, (s,a), the Q-learning function updates the current estimate based on new information received from recent actions, r, and discounted long term reward. In general, an action is selected from a given state, the reward outcome is observed and recorded, and the value function updated. The value associated with each action is used during each visit to a particular state to determine which action should be chosen using the Boltzmann distribution, shown below.

Pi =

e

) Q ( σ , a )i τ

∑e

) Q (σ , a ) j τ

(1.1)

j

The Boltzmann distribution uses the temperature term, τ, to control the level of exploration and exploitation. A high temperature translates into exploratory behavior, a low temperature results in greedy behavior. 2.3 Cognitive Social Simulation Cognitive social simulations are particularly well-suited for defense applications which typically call for analysis of a military force’s performance while operating as part of a complex conflict ecosystem [6]. Agent based models have been applied to the military domain previously [7], but the use of cognitive social simulation, enabled by cognitive architectures, in this application area is relatively new [8]. The relevancy of these tools is particularly highlighted by the nature and objectives of the current conflicts, where the population of the conflict area is seen as the center of gravity of the operation [9]. Gaining an understanding of potential means of transitioning the social system in these countries from an unstable to a stable state provides a challenge to leaders at all levels of the military and civilian organizations involved. The representation of trust and its impact on the effectiveness of information operations is required in order to provide greater insight into the population within the area of interest.

3 Approach This section provides an overview of a general model of learned trust for agents in cognitive social simulations, an introduction to the test bed environment, and a description of the proof of principle implementation in Python 2.6.

304

S.S. Pollock, J.K. Alt, and C.J. Darken

3.1 General Model This general model is a turn-based simulation in which the agents and their relationships are represented by a single network graph with the agents as the nodes and their social relationships as the edges, weighted by the value of the relationship. To be specific the network graph is a combination of two similar graphs with identical nodes, but in one the edges are bidirectional and represent a base constant value and the second one in place of each bidirectional edge contains a pair of directed edges representing the agent’s individual variable contribution to the strength in the relationship. In this way, the edge weights have both a static and dynamic component. The static is based entirely on the concept of homophily ( ), that like persons associate more frequently, utilizing a simple Euclidean distance calculation and as stated above is set as a constant [10]. The dynamic portion is completely under the control of the agents involved ( ). In the case of k total agents, each agent has k-1 choices of who to spend time with and based on that emphasis will receive some unknown reward from these relationships. For every simulation round, each agent will choose to increase, decrease or maintain their contribution to their relationships with the other agents. This contribution can be seen like a fraction of time spent with the others in that it is represented as a floating point number from 0.0 to 1.0 and such that the sum of all these components (i.e. the sum of all edge weights leaving the node representing the agent) always sums to 1.0. At the end of each turn the agent is rewarded based on the strength of their relationships. This reward takes the following form: · min , (1) The equation uses the minimum of the variable contributions from each agent. In this way it more accurately can be said that the variable portion is the fraction of time that the agent would like to spend with the other agent and therefore using the min also provides a small penalty for those who place more emphasis on a relationship than the recipient. The result of this basic model is the development of a simple dynamic social network. The network tends to become highly centralized around 1 or 2 agents. In particular, in runs consisting of 50 agents, the final network graph consisted of nearly every agent with a strong connection to a single central agent with no other connections present. In order to mitigate this affect it was necessary to add a second order factor in the reward calculation for the agents. Continuing along the analogy that the emphasis represents a fraction of time desired to be spent with the other agents, then it is natural to extend this and allow for groups of more than 2 agents to spend time together. In other words if two agents have a strong relationship and also each have a strong relationship to the same third agent, then all three agents should receive an additional reward for this. The second order reward factors are based on the same reward function as used in the first order above. In this case, the reward is divided by a distribution factor and subsequently squared. For the case of agents A and B as above, but this time having a common friend in agent C, the additionally reward looks as below: 2

min

· min

,

,

· min

,

/

(2)

Representing Trust in Cognitive Social Simulations

305

The closeness centrality of the network is highly sensitive to the distribution factor and is discussed in detail in section 4. Once the second order terms are added similar network properties to what we would expect to see in real social situations emerge; namely subdivisions into clique’s, pairings and the exclusion of certain individuals. It is obvious that this feature is not intended to actually model the internal and external processes that form human social networks; rather it is simply building the stage on which to build future work. 3.2 Prototype Trust Implementation Each agent has a simple belief structure consisting of a finite set of beliefs, five in this case, represented by a single floating point number from 0.0 to 1.0.These beliefs combine in simple linear combinations to provide issue-stances, one in this case, also as a floating point from 0.0 to 1.0. During each turn of the simulation, following an initial stabilization period, set initially to 1000 rounds for a simulation of 15 agents, the agent will choose a topic based on a probabilistic Boltzmann distribution, and discuss this topic with its k nearest neighbors. Other than the k nearest there any neighbor above a specified threshold will always receive a communication, while neighbors below another threshold will never receive a communication. Initially, the communications consist of each agent telling his closest neighbors what his value is on a selected belief. The receiving agents then use a reinforcement learning algorithm to determine whether or not belief revision is merited. In order to utilize reinforcement learning it is necessary to define some concept of a reward that the agent will receive based on their beliefs and therefore directly related to their trust and belief revision mechanisms. Our inspiration for a reward model comes from Game Theory and is called the Public Commodity Game [3]. In this game, each agent has an option to contribute to a public pot of money each round or to opt out. Following the round the money in the pot is multiplied by some amount (in this case 3.0) and then redistributed to each agent regardless of contribution. As a slight variation of this, the agents are allowed to decide an amount to contribute rather than to choose from full or no contribution. In the current model, agents are given 1.0 possible units to play such that an agent that contributes nothing is guaranteed a reward of at least 1.0 for opting out and an unknown reward ranging from nearly 0 to 3.0 for full contribution. Game theory tells us that without cooperation the expected equilibrium for rational players would be exactly 0.0 contributions from all agents, in other words all agents take the guaranteed 1.0 and opt-out of the public commodity all together [3]. It is likely the case that some people would always contribute at least some small amount irrespective of their losses. In order to achieve this, “Faith in the Public Commodity” is the issue-stance and is used to directly control the level of their contribution to the public commodity. During each simulation round, agents communicate with one another and attempt to bring other agents closer to their beliefs. Despite the fact that a 1.0 contribution from all agents is the most mutually beneficial strategy, it is not a stable equilibrium as a single person could quickly realize that decreasing their contribution will increase their total revenue and would be easily reinforced by learning algorithms. What is seen is the agents benefit the most from a strategy of trying to make all the other agents contribute at a level consistent with their issue strength.

306

S.S. Pollock, J.K. Alt, and C.J. Darken

Now that there is a concrete idea of reward in this model, it is possible to begin applying a simple model of belief revision. Essentially, the agent will use a reinforcement learning algorithm where the state space is represented by sender-topic pairs and the action will be to dismiss the communication or to apply a revision, based on their level of trust in the information, resulting in an update of the fraction of time they desire to spend with the sender. Without any further weighting factors or limiting algorithms, the expected result of this simulation is that the beliefs of all agents will approach a stable equilibrium in which all beliefs are the same.

4 Experimentation and Analysis This section provides a comparison of model results from several attempts to moderate belief revision within the model. 4.1 Experimentation of Second Order Factors in the Base Social Network The first task was to evaluate the effects of the strength of the second order terms in the base social network reward functions. The distribution factor was varied and showed a fairly steep “S” curve that was centered between D = 14 to D = 24.

Fig. 1. Closeness centrality versus distribution factor

The distribution factor in affect allows fine tuning of closeness centrality in the base model in order to fit it to any particular purpose. There are several widely varying sources on what a real human social network should look like in terms of closeness centrality that range from 0.20 to 0.60. Therefore, for the purposes of the remainder of this initial experimentation D = 18.4 is used in order to target the fractional closeness centrality to around 0.30. The exact nature of these values is irrelevant for this initial model and only serves as a baseline for further work. 4.2 Belief Revision Experimentation As discussed previously it is necessary to implement within the model a method for moderating how the agents can revise their beliefs. Initial methods for doing this are, at least for now, centered on a penalty for changing beliefs.

Representing Trust in Cognitive Social Simulations

1.0 –

307

3.0

Where the Belief Variance is a simple Euclidean distance measure from the agents current beliefs to what they started with at the beginning of the simulation. What is surprising is that when F is varied, there is no marked difference in the outcome of the simulation from a purely statistical point of view. What is seen however is in the Public Commodity play over time.

Fig. 2. Public commodity game play versus norm penalty

The higher the factor F becomes, the more unstable the Public Commodity game is. In other words, with a small norm penalty the agents will tend to find a stable equilibrium and remain there with fairly significant stability. As F is increased, the stability is decreased. The intriguing thing is that this behavior appears to be similar to actual human interaction. For example, if we look at a society there is a sense of a norm although it will really change over time it will remain fairly constant over small enough time periods. In this society there will be people or factions that challenge the social norm causing brief unstable equilibrium away from the norm that seem to return to the social norm after some time. More investigation into this phenomenon is currently underway. Once there is a satisfactory explanation for this behavior it may be possible, just as in the second order parameter, to tune the magnitude of the norm penalty in a way that is unique to the society being modeled.

5 Conclusions and Future Work The results of the initial work are promising in that it is seen that there are factors within this simple first draft model that allow the social network to form and evolve based purely on the agents cognitive processes however the nature of the network can be tuned to match a desired case. This initial effort to model trust shows a lot of promise and merits further investigation. The next step in this process will be to utilize this algorithm within an existing social simulation and evaluating the effects. The next generation of this model is planned to include a much more complicated belief structure that will include several metacognitive elements giving the agents some control over belief revision. There will also be included within the belief structure mental models of the other agents and the environment so that a perceived closeness in beliefs of the other agents will play directly into trust.

308

S.S. Pollock, J.K. Alt, and C.J. Darken

References [1] Anderson, J.: Cognitive Psychology and Its Implications, 6th edn. Worth Publishers, New York (2005) [2] Sun, R.: Cognitive social simulation incorporating cognitive architectures. IEEE Intelligent Systems, 33–39 (2007) [3] Owen, G.: Game Theory. Academic Press, San Diego (1995) [4] McKnight, D., Chervany, N.: The Meanings of Trust. University of Minnesota (1996) [5] Sutton, R., Barto, A.: Reinforcement Learning: An Introduction. MIT Press, Cambridge (1998) [6] Kilcullen, D.J.: Three Pillars of Counterinsurgency. In: US Government Counterinsurgency Conference [7] Cioppa, T.M., Lucas, T.W., Sanchez, S.M.: Military applications of agent-based simulations. In: Proceedings of the 36th Conference on Winter Simulation, p. 180 (2004) [8] National Research Council, Behavioral Modeling and Simulation: From Individuals to Societies. National Academies Press, Washington DC (2008) [9] C. O. S. W. Department of the Army, Counterinsurgency. U.S. Army (2006) [10] McPherson, M., Smith-Lovin, L., Cook, J.M.: Birds of a feather: Homophily in social networks. Annual Review of Sociology 27(1), 415–444 (2001)

Temporal Visualization of Social Network Dynamics: Prototypes for Nation of Neighbors Jae-wook Ahn1,2 , Meirav Taieb-Maimon2,3 , Awalin Sopan1,2 , Catherine Plaisant2 , and Ben Shneiderman1,2 1

2

Department of Computer Science Human-Computer Interaction Lab, University of Maryland, College Park, MD, USA {jahn,awalin,plaisant,ben}@cs.umd.edu 3 Ben-Gurion University of the Negev, Beer-Sheva, Israel [email protected]

Abstract. Information visualization is a powerful tool for analyzing the dynamic nature of social communities. Using Nation of Neighbors community network as a testbed, we propose five principles of implementing temporal visualizations for social networks and present two research prototypes: NodeXL and TempoVis. Three different states are defined in order to visualize the temporal changes of social networks. We designed the prototypes to show the benefits of the proposed ideas by letting users interactively explore temporal changes of social networks. Keywords: Social Network Analysis, Information Visualization, Social Dynamics, User Interface, Temporal Evolution.

1

Introduction

Information visualization is a powerful tool for analyzing complex social networks so as to provide users with actionable insights even against large amount of data. Numerous social network analysis (SNA) methods have been coupled with visualization to uncover influential actors, find helpful bridging people, and identify destructive spammers. There are many research tools and a growing number of commercial software packages, some designed for innovative large scale data analysis, while others deal with common business intelligence needs. However, few approaches or tools sufficiently address the problem of how to analyze the social dynamics of change over time visually. Communities are not static. Like living organisms, they evolve because of cultural, environmental, economic, or political trends, external interventions, or unexpected events [4]. Technological developments have also had strong impacts on social changes, a phenomenon that has become influential with the arrival of mobile communications devices and social networking services. Nation of Neighbors (http://www.nationofneighbors.com) (NoN) is a new web-based community network that enables neighbors to share local crime, suspicious activity, and other community concerns in real time. The NoN developers J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, pp. 309–316, 2011. c Springer-Verlag Berlin Heidelberg 2011 

310

J.-w. Ahn et al.

have achieved an admirable success with “Watch Jefferson County” that empowers community members to maintain the security in their local neighborhoods. It began in Jefferson County, WV, but the NoN team is expanding their efforts across the U.S. in many communities. We are collaborating with them to provide appropriate tools that can help community managers explore and analyze the social dynamics embedded in their social networks. This paper introduces our first efforts to visualize the temporal social dynamics on top of the NoN testbed. Due to the nature of the social dynamics, we need a method to interactively compare the time slices of social networks. In previous attempts, we combined statistics and the visualization [12] for interactive network analysis and to support multiple network comparisons using tabular representations [3]. The current work integrates the best of these two approaches by using the graphical representations and statistics to enable researchers and community managers to compare networks over time.

2

Related Studies

The importance of capturing the network change over time has been the origin of previous attempts of temporal network visualizations. Two common approaches were used. The first approach plots network summary statistics as line graphs over time (e.g. [1,7]). Due to its advantage in detecting the increase and decrease of certain statistics, many systems adopted this idea. In the VAST 2008 mini challenge 3:“Cell Phone Calls” [8] various teams (e.g. SocialDynamicsVis and Prajna teams) used similar approaches such as line graphs to characterize changes of the cell phone call social network over ten days. SocialAction integrated a time-based interactive stacked graph to display a network property such as degree. Each stack represented a node and the thickness of each stack on a given day represented its degree. The interactive visualization facilitated the exploration of the change in the degree of each node over time. The second approach is to examine separate images of the network at each point in time. Powell et al. [13] used such a snapshot approach to show network changes over time. They distinguished new arrivals with triangles and incumbent with circles. They also used size encoding for degree; a shortcoming of such size and shape encoding might be that the new arrivals having small degree are very small and sometimes the triangles and circles are too small to be distinguishable. Furthermore, in snapshot view we cannot compare a network easily in two timeslots as the layout algorithms can dynamically change the node positions. Durant et al. [2] presented a method for assessing responses and identifying the evolution stages of an online discussion board. Their snapshot-based network visualizations could show different node positions between time. Recent work offered improvements in static visualizations by applying dynamic visualization such as network “movies.” Moody [11] distinguished flipbook style movies where node-positions were fixed but connectivity formation was captured and dynamic movies where node-positions changed over time. Condor (or TeCFlow) [5,6] introduced sliding time frames and animated the network

Temporal Visualization of Social Network Dynamics

311

changes inside a specific time frame with/without the history before that. However, approaches which use animations might be distracting for users to track changes in the network. Specifically, it might be difficult to track new nodes when the position of the nodes and edges keep changing. Human eyes can miss minute but important changes in visual spaces. For example, users can only notice the overall growth of an entire network while missing that the inclusion of a certain node is responsible for a significant change. A possible suggestion can be found in Trier’s work [14] where he used node degree as a measure of inertia so the high degree nodes move less across time and made the dynamic movie less distracting. C-Group [10] visualized the affiliation of focal pairs and dynamically defined groups over time using static, dynamic layouts, or animations. It fixed the focal pair positions and could minimize distraction in the animation mode. In the present study, we suggest an approach that keeps the nodes from earlier time slots in fixed positions but also accentuates new arriving nodes and links using color coding while using intensity as a measure of aging. This overcomes the distraction caused by dynamic fast paced network movies and also fulfills the need of identifying the deletion and addition of nodes.

3

Social Network Dynamics Visualization – Ideas

Our main concern about the temporal dynamic visualization for social networks is how to show and compare the changes over time. We established the following principles for this purpose. 1. 2. 3. 4. 5.

Static (unchanged) parts of a graph should remain fixed to avoid distraction. Changes should be visually manifest for easy comparisons. Comparisons across time should be easily and interactively explorable. Temporal change of the node and edge attributes should be discoverable. Temporal change of a sub-graph and its attributes should be discoverable.

Some approaches do not fix the locations of the static elements, but since users’ main interests are about what is changing, fixing the locations of the unchanged nodes and edges is advantageous. For the changing components, we defined three states: (1) addition, (2) removal, and (3) aging. In social networks, new relationships can be added or removed to/from the entire network. At the same time, existing relationships can become inactive after they were added to the network. Especially, there are instances where the removal of the relationships is not quite clear and older activities are just becoming inactive. For example, a reference once made in a citation network cannot be removed, even though it can be less important as time passes. We tried to distinguish these three states in our network visualization. Finally, users need to easily compare the network at any pair of time points.

4

Social Network Dynamic Visualization Systems

Based on these principles, we built two example systems: (1) by using NodeXL and (2) by adopting a time-slider for temporal exploration. NodeXL [9] is an

312

J.-w. Ahn et al.

Fig. 1. Visualizing the NoN invitation network dynamics using NodeXL. Nodes and edges mean users and invitations respectively. Red edges are the ones in September 2010. Using the dynamic filter (right), it can filter out the edges before August and emphasize more recent invitations (8/1/2010 – 9/14/2010, the edges in the upper part).

extension to Microsoft Excel 2007/2010 that makes the SNA flexible and interactive. The node-edge connectivity data stored in Excel worksheets are directly converted to graph visualizations. Popular SNA statistics can be calculated and then be mapped to visual attributes. It is highly configurable by users through the common Excel operations and the NodeXL metrics, visual properties, clustering, and layout control panels. The second system called TempoVis was built as a standalone application prototype, in order to implement a time-based network exploration more interactively. We used two datasets provided by NoN: user invitation and forum conversation networks. The former contained information about who invited whom to the NoN service (July 2009 – September 2010). The latter showed which user created a thread and who replied to whom in the NoN forums (November 2005 – August 2010) with timestamps. 4.1

Spreadsheet-Based Approach Using NodeXL

Fig. 1 shows an example of temporal visualization using NodeXL. The worksheet stores the node (vertex) and edge information. The graph theory statistics such as centrality are calculated for each node too (for example, Vertex 1831’s Betweenness Centrality=3444.0). In the right-hand side panel, the network is visualized using the well-known Fruchterman-Reingold force-directed layout algorithm (the layout was computed from the latest time point). It visualizes the invitation network of the NoN members accumulated for 14 months. The overall network layout does not change by the dynamics of the invitation network over time. However, the temporal dynamics in the individual invitation level are represented clearly. The red edges (arrowed) in Fig. 1 indicates that

Temporal Visualization of Social Network Dynamics

313

they were added in a specific time point, September 2010. At the same time, the nodes and edges created earlier than a specific time (or removed) can be painted in different colors or dynamically filtered out as in Fig. 1 in order to implement the three states (Section 3). The graph elements are tightly linked to the worksheet, so that users can examine specific attributes of any nodes, edges, or subgraphs interactively while they are switching between the worksheet and the visualization. However, despite the very rich and flexible feature set that helps users to compare the data and the visual representation interactively, easy exploration across multiple time points is yet to be added to NodeXL. Even the dynamic filter requires some manual endeavors. Therefore, we built the next prototype that focused more on easier time-based exploration. 4.2

TempoVis: Adding Time-Based Interactive Exploration

TempoVis (Fig. 2) is a standalone prototype that shares the same principles but we designed it to maximize the ability to explore the time interactively. It is equipped with a time-slider with which users can navigate through different time points (months in this example). They can move the slider thumb left and right, to view the different monthly states. Users can navigate back and forth through

Fig. 2. TempoVis: Navigating through time using a time-slider interface (NoN conversation network). The nodes and the edges mean users and their conversations respectively. The red edges indicate the conversations of a specific month (June 2010) set by using the slider blow. The big low-intensity cluster (center below) was introduced earlier than the high intensity cluster (center top) that was added recently. Blue edges mean user selections. Node size is proportional to graph statistics dynamically calculated at each time point.

314

J.-w. Ahn et al.

time to study the growth of the network. On top of the slider, a histogram shows the frequency of the activities. This arrangement enables users to see the overview of the activities and compare the changes over time during navigations. TempoVis uses the force-directed layout and the red colored nodes/edges indicate new activities of the current month. The slider is set to June 2010 in Fig. 2 and the red edges and nodes represent the conversations added in that month. The gray items are the conversations that took place before June 2010. Fig. 3 (a) and (b) show how users can compare adjacent months using the time-slider and visualization. By switching (a) and (b) using the slider, they can easily discover how many new conversations (three) were added. Whereas (a) and (b) cannot give information about removed or aging conversations, Fig. 3 (b’) distinguishes the current month’s activities and the other older conversations by gradually decreasing the intensity as they age. We can see that the conversations added in the previous month (red in (a)) are painted in higher intensities while some far older nodes/edges are barely visible. This effect gives users information about which activities were more recent and which activities took place long before. Zhang et al. [15] used a similar approach to represent the age of the edges in an invitation network but they did not support the dynamic update of the edge intensities by navigation. A sigmoid function was used in order to implement the smooth drop of the color intensity by time. Even though this network visualization provides users with information about the temporal change about the entire network, they may also be interested in changes in part of a graph. Therefore, we provided a marquee-selection tool that lets users select any part of the graph. For example, in Fig. 2, the user was interested in a small group of conversations (highlighted in blue by the user). S/he could marquee select them and then the information about the selected edges were dynamically overlaid on the histogram, showing that the conversations were done in March 2010 and the frequency was 9.

Fig. 3. Comparing February (a) and March 2007 (b and b’). Older activities can be painted in lower intensities (b’) to show their age.

5

Discussions and Conclusions

This paper introduced our efforts to visualize the temporal dynamics of social networks. Five principles for implementing visualizations that support interactive exploration over time were proposed. We presented two prototypes according

Temporal Visualization of Social Network Dynamics

315

to these principles and showed two datasets from the Nation of Neighbors social networking service: one utilizing the NodeXL extension to Microsoft Excel and TempoVis as a standalone application. The former approach is very flexible and easy to access because it is based on the well-known spreadsheet application and highly configurable. TempoVis was aimed to test the potential of a user interface element that could facilitate temporal navigation – a time-slider. We expect the NodeXL-based idea will be preferred by the users who are proficient with the data manipulation using spread sheets. TempoVis will be useful for the tasks where more exploratory and interactive temporal comparisons are required. In order to evaluate our prototypes and understand their usefulness from a network manager’s point of view, we consulted with Art Hanson, the NoN lead designer. He said: “The temporal visualizations would help focus limited resources on community groups who could best benefit from assistance, spot emerging trends, and identify the most influential users. The tools would also help NoN visualize the network effects resulting from changes to the NoN system and achieve the desired effect quickly with each new feature or design iteration.” He could easily identify which group of NoN users showed different activities at a specific time point and use that information to manage and evaluate the community service. It was achievable by combining the time and the network information of the neighbors provided by our temporal visualization tools. We are planning more extensive user studies to validate the usefulness of our five principles and its implementations in the two visualization platforms. The NoN team suggested to add the ability to distinguish node types (e.g. post and report) to the visualizations and to add the community group or location/distance information, so that they could link the temporal dynamics to community group activities or communications types. At the same time, we will also investigate how to locate sub-graphs and analyze their temporal dynamics. While fixing the locations of stable nodes/edges (principle 1) is useful, we are also planning to add a user option to dynamically redraw the graph layouts. The tasks such as observing large growth and change of clusters will require this feature. We are also considering to adopt multiple network views of user-selected time points in order to let users directly compare multiple time points. Enriching the visualization with additional attributes such as discussion topics will help differentiate various activities too. After more evaluation with community managers and network analysts we will embed the most valuable features into the NodeXL platform, while continuing to explore novel features in TempoVis. Better understanding of network dynamics is necessary so that community managers can discover the triggers for growth, decay, active use, or malicious activities. Then they can develop strategies to promote growth and prevent decay and destructive interference. Social media have great potential for benefits in key domains such as community safety, disaster response, health/wellness, and energy sustainability. Better understanding of how to manage them effectively raises the potential for more successful outcomes.

316

J.-w. Ahn et al.

Acknowledgements. We appreciate the support of Microsoft External Research for the development of the free, open source NodeXL (http://www. codeplex.com/nodexl) and for National Science Foundation grant IIS0968521, Supporting a Nation of Neighbors with Community Analysis Visualization Environments, plus National Science Foundation-Computing Research Association Computing Innovation Fellow, Postdoctoral Research Grant for Jae-wook Ahn. We appreciate the assistance of Art Hanson and the NodeXL team.

References 1. Doreian, P., Kapuscinski, R., Krackhardt, D., Szczypula, J.: A brief history of balance through time. The Journal of Mathematical Sociology 21(1), 113–131 (1996) 2. Durant, K., McCray, A., Safran, C.: Modeling the temporal evolution of an online cancer forum. In: 1st ACM International Health Informatics Symposium (IHI 2010), pp. 356–365 (November 2010) 3. Freire, M., Plaisant, C., Shneiderman, B., Golbeck, J.: Manynets: an interface for multiple network analysis and visualization. In: Proceeding of the 28th Annual SIGCHI Conference on Human Factors in Computing Systems, pp. 213–222 (2010) 4. Giddens, A.: Sociology. Polity, Cambridge (2006) 5. Gloor, P., Laubacher, R., Zhao, Y., Dynes, S.: Temporal visualization and analysis of social networks. In: NAACSOS Conference, June, pp. 27–29 (2004) 6. Gloor, P., Zhao, Y.: Analyzing actors and their discussion topics by semantic social network analysis. In: Tenth International Conference on Information Visualization, IV 2006, pp. 130–135. IEEE, Los Alamitos (2006) 7. Gould, R.: The origins of status hierarchies: A formal theory and empirical test. American Journal of Sociology 107(5), 1143–1178 (2002) 8. Grinstein, G., Plaisant, C., Laskowski, S., O’connell, T., Scholtz, J., Whiting, M.: VAST 2008 Challenge: Introducing mini-challenges. In: IEEE Symposium on Visual Analytics Science and Technology, VAST 2008, pp. 195–196. IEEE, Los Alamitos (2008) 9. Hansen, D.L., Shneiderman, B., Smith, M.A.: Analyzing Social Media Networks with NodeXL: Insights from a Connected World. Morgan Kaufmann, San Francisco (2010) 10. Kang, H., Getoor, L., Singh, L.: Visual analysis of dynamic group membership in temporal social networks. ACM SIGKDD Explorations Newsletter – Special Issue on Visual Analytics 9(2), 13–21 (2007) 11. Moody, J., McFarland, D., Bender-deMoll, S.: Dynamic network visualization. American Journal of Sociology 110(4), 1206–1241 (2005) 12. Perer, A., Shneiderman, B.: Integrating statistics and visualization: case studies of gaining clarity during exploratory data analysis. In: Proceeding of the 26th Annual SIGCHI Conference on Human Factors in Computing Systems, pp. 265–274 (2008) 13. Powell, W., White, D., Koput, K., Owen-Smith, J.: Network Dynamics and Field Evolution: The Growth of Interorganizational Collaboration in the Life Sciences. American Journal of Sociology 110(4), 1132–1205 (2005) 14. Trier, M.: Towards dynamic visualization for understanding evolution of digital communication networks. Information Systems Research 19(3), 335–350 (2008) 15. Zhang, J., Qu, Y., Hansen, D.: Modeling user acceptance of internal microblogging at work. In: CHI 2010 Workshop on Microblogging (2010)

Toward Predicting Popularity of Social Marketing Messages Bei Yu1, Miao Chen1, and Linchi Kwok2 2

1 School of Information Studies College of Human Ecology, Syracuse University {byu,mchen14,lkwok}@syr.edu

Abstract. Popularity of social marketing messages indicates the effectiveness of the corresponding marketing strategies. This research aims to discover the characteristics of social marketing messages that contribute to different level of popularity. Using messages posted by a sample of restaurants on Facebook as a case study, we measured the message popularity by the number of “likes” voted by fans, and examined the relationship between the message popularity and two properties of the messages: (1) content, and (2) media type. Combining a number of text mining and statistics methods, we have discovered some interesting patterns correlated to “more popular” and “less popular” social marketing messages. This work lays foundation for building computational models to predict the popularity of social marketing messages in the future. Keywords: text categorization, marketing, social media, prediction, media type.

1 Introduction Social media has become a major influential factor in consumer behaviors, such as increasing awareness, sharing information, forming opinions and attitudes, purchasing, and evaluating post-purchasing experience [8]. Among those social media tools that are available for businesses, Facebook has become one of the most important media for B2C and C2C communications. Today many websites embed Facebook’s “Liked” buttons, which not only allows consumers to exchange information within their Facebook network in a much faster speed but also pushes companies to pay more attention to Facebook fans’ reactions to their messages sent through social media. Facebook fans’ endorsement of a company’s messages could be an important indicator of how effectiveness a company’s social marketing strategies are. This research1 aims to examine the effectiveness of different social marketing strategies. In social media such as Facebook, the long-term effectiveness of marketing strategies, such as the impact on consumer behavior, might not be directly measurable. However, Facebook provides some measures that indicate the message popularity, which is the level of attention that fans paid to the messages, such as the number of 1

This research project is supported by the Harrah Hospitality Research Center Grant Award Program from University of Nevada, Las Vegas.

J. Salerno et al. (Eds.): SBP 2011, LNCS 6589, pp. 317–324, 2011. © Springer-Verlag Berlin Heidelberg 2011

318

B. Yu, M. Chen, and L. Kwok

“likes”, the number of comments, and the overall tone of the comments. To some extent the popularity of a message indicates the effectiveness of the marketing strategy embedded in this message. Therefore in this research we aim to discover the characteristics of social marketing messages that affect their popularity. We focus on the number of “likes” they attract, and leave the comment analysis for future work. Using the number of “likes” to measure the popularity of a message, we aim to answer the following research question: what characteristics of the messages contribute to their popularity? We examine the characteristics of messages from two perspectives: (1) content, and (2) media type. This article is organized as follows. Section 2 describes the data preparation process, including data sampling, downloading, and cleaning. Section 3 describes the thematic analysis of the message content. Section 4 describes an approach to predicting popularity based on the message content. Section 5 describes the correlation analysis between media type and message popularity. Section 6 concludes with limitations and future work.

2 Data Sampling, Downloading, and Cleaning In this section we describe the data sampling, downloading, and cleaning process. We choose restaurant industry as a case study for the following reasons. Restaurant business has always been a big component of the tourism and hospitality industry, where intangible and perishable products are produced and consumed. Due to the experiential nature of these service products, consumers often seek information from online reviews and travel blogs before making purchasing decisions [10]. Restaurants understand the importance of communicating the “right” messages with their existing and potential customers over the Internet. Recent business reports also showed that restaurant entrepreneurs can substantially draw business and increase sales with Twitters and Facebook [11]. Many restaurants have been actively engaged with their customers on Facebook, which makes restaurant industry a great example for examining companies’ social marketing messages. This research adopted a criterion-based, mixed-purpose sampling procedure and selected twelve restaurants’ Facebook pages in September 2010 based on the following criteria and steps: (1) we examined the top 20 restaurant chains with the highest sales volume [7] and selected all quick service restaurant chains with more than one million Facebook fans (eight in total) and the top three casual dining restaurant chains with the most Facebook fans; (2) we examined the top 20 independent restaurants and selected the top two with the most Facebook fans. We should have selected 13 restaurants in total, but eventually included 12 because KFC’s Facebook page couldn’t be correctly identified. We then identified each restaurant’s unique Facebook ID and retrieved all posts using Facebook FQL. We retrieved three fields: message body, message’s media type (“status”, “link”, “video”, or “photo”), and the number of people who have clicked the “likes” button on this message. On October 5th we retrieved a total of 675 messages posted by the above 12 restaurants. See Table 1 for the number of posts collected for each restaurant. So far Facebook FQL only returns posts in recent 30 days or 50 posts, whichever is greater. We plan to continue our data collection on a regular basis for the next six months and re-examine our analysis on the larger data set in the future.

Toward Predicting Popularity of Social Marketing Messages

319

Because the number of fans for each restaurant varies significantly, messages from different restaurants with same number of “likes” could indicate significantly different level of popularity. For example, the most popular message from independent restaurant Carmines attracted only 93 likes, while the least popular message from Starbucks attracted 2,250 likes. To make the numbers of “likes” comparable among restaurants we normalized the numbers of “likes” for each restaurant using z-score. The normalized numbers of “likes” are distributed in the range (-3, 2). Table 1. Descriptions of the 12 restaurants in this study Restaurant Name Quick Service (7) McDonald’s Starbucks Subway Pizza Hut Taco Bell Dunkin’s Donuts Chick-fil-A Casual Dining (3) Chilli’s Grill & Bar Olive Garden Outback Steakhouse Independents (2) Joe’s Stone Crab Carmine’s

Facebook ID 50245567013 22092443056 18708563279 6053772414 18595834696 6979393237 21543405100

No. of FB Fans 3,244,673 14,605,410 2,218,422 1,473,274 2,842,184 2,078,568 3,106,304

No. of FB posts 39 58 58 63 102 45 73

106027143213 18618646804 48543634386

470,994 544,278 883,670

51 41 40

404077308008 99057540626

1,628 5,190

54 51

3 Thematic Analysis Identifying social marketing strategies is difficult without an established typology. However we may identify the common themes of the messages as indicators of restaurants’ favorite marketing strategies. For example, if “donation” occurs frequently we can infer that charity is a commonly used marketing strategy. In order to identify the main themes we first applied Latent Dirichlet Allocation (LDA) algorithms [1] to all messages and identified ten main themes. However, the resulted themes were not as meaningful probably because LDA has proved to be more effective on larger data set. We then turned to basic keyword analysis to identify the most common content words (nouns, verbs, adjectives, and adverbs) that each restaurant used as a representation of its marketing strategies. Because of the important role of pronouns in communication [2], we also counted pronouns in this study. During the keyword analysis process we (1) gathered all messages posted by each restaurant as an individual text collection; (2) tokenized the messages and computed each token’s overall term frequency (how many times it occurs in this text collection); (3) manually picked the ten most frequent pronouns and content words that consist of the “keyword profile” of each restaurant. The resulted keyword profiles are presented in Table 2.

320

B. Yu, M. Chen, and L. Kwok

The keyword profiles in Table 2 depict a common communicative style signified by prevailing use of first and second person pronouns like “you”, “your”, “we”, and “our”. These profiles also uncover two common marketing strategies. The first is to promote key concepts as part of their public images. For example, McDonald’s promoted “family time”, Subway emphasized “fresh” food and being “fit”, and Starbucks focused on the concept of “enjoy”. The second common strategy is to run advertisement campaigns. Many of the keywords other than pronouns involve the major advertising campaigns, especially contests. Almost all restaurants launched some contests, for example “Super Delicious Ingredient Force” by TacoBell, “Ultimate Dream Giveaway” by Pizza Hut, “Create-A-Pepper” by Chili’s, “Never Ending Pasta Bowl” by OliveGarden, “Free Outback For a Year” by Outback, “Eat Mor Chickin Cowz” by Chick-fil-A, “Fan Of The Week” by Dunkin’s Donuts, etc. Table 2. Keyword profiles for each restaurant Restaurant Quick Service (7) McDonald’s Starbucks Subway Pizza Hut Taco Bell Dunkin’s Donuts Chick-fil-A Casual Dining (3) Chilli’s Grill & Bar Olive Garden Outback Steakhouse Independents (2) Joe’s Stone Crab Carmine’s

Keywords you, your, games, all, we, time, Olympic, angus, more, howdoyoumcnugget you, we, your, coffee, our, free, us, about, enjoy, more your, you, favorite, fresh, footlong, today, chicken, fit, buffalo, breakfast your, pizza, you, our, today, we, order, free, new, dream you, your, who, force, night, sdif, fourthmeal, super, chicken, delicious you, FOTW, our, coolatta, week, DD, next, your, bit, submit cowz, you, mor, your, cow, day, facebook, our, like, appreciation we, you, fans, our, your, pepper, us, winner, St. Jude, gift you, our, your, pasta, new, we, ending, bowl, never, contest bit, free, your, our, year, you, win, check, visit, chance special, week, our, dinner, crab, jazz, summer, we, you, sauce your, our, we, family, memories, my, Washington DC, us, you, dinner

Did the key concepts resonate among fans? Did these contests really catch the eyes of fans? Or more broadly which strategies are more effective? By measuring the effectiveness with the popularity of relevant messages, we translated this question to the question of which messages are more popular than others, and then adopted an exploratory approach of using text categorization and feature ranking methods to discover the relationship between message content and popularity.

4 Content-Based Message Popularity Prediction We grouped all messages into two categories: the “more popular” category for messages with positive normalized numbers of “likes” (above average), and the “less popular” category for messages with negative normalized numbers of “likes” (below average). After this process, the messages are separated into two categories: 430 messages in the “less popular” category and 207 messages in the “more popular” category. To build a sharp and balanced classifier we formed a new data set “400

Toward Predicting Popularity of Social Marketing Messages

321

messages” which consists of 200 most popular and 200 least popular messages. Each message is tokenized and converted to a word vector using bag-of-word representation. To focus on common words we excluded words that occurred in fewer than three messages and resulted in a vocabulary of 568 words. Reducing the vocabulary size also resulted in 19 empty vectors in the “400 messages” data set, which eventually has 381 (192 “more popular”, and 189 “less popular”) messages. We then apply text categorization algorithms to examine the extent to which the “more popular” and “less popular” categories are separable. We employed two widely used text categorization algorithms Support Vector Machines (SVMs) and naïve Bayes (NB) to train and evaluate the content-based popularity prediction models. SVMs are among the best methods for text categorization [6] and feature selection [4]. SVMs favor discriminative features with broad document coverage, and therefore reduce the risk of over-fitting. Studies have shown that the 10% most discriminative features identified by SVM classifiers usually work as well as, or even better than the entire original feature set [12]. In this study we used the SVM-light package [5] with default parameter setting. We compared the performance of four variations of SVM models: SVM-bool with word presence or absence (one or zero) as feature values; SVM-tf with raw word frequencies; SVM-ntf with word frequencies normalized by document length; and SVM-tfidf with word frequencies normalized by document frequencies. Naïve Bayes (NB) is another commonly used text categorization method, simple but effective even though the text data often violate its assumption that feature values are conditionally independent given the target value [3]. We compared two variations of naïve Bayes implementations: NB-bool, the multi-variate Bernoulli model that uses word presence and absence as feature values, and NB-tf, the multinomial model that uses raw word frequencies. We implemented the two naïve Bayes variations. We used leave-one-out cross validation to evaluate the SVMs and NB classifiers on both “all messages” and “400 messages” data sets. We used the majority vote as the baseline for comparison. The results in Table 3 show that all classifiers beat the majority vote on the “400 messages” data set by wide margins, and that SVM-ntf and SVM-bool both achieved accuracy over 75%. In contrast the cross validation accuracies on the “all messages” data set are less encouraging in that they are merely comparable to the majority vote. The good result on “400 messages” demonstrates that the “more popular” and “less popular” messages are fairly separable based on message content. The poor result on “all messages” also supports this notion in that the messages with medium level of popularity proved to be difficult to categorize. Table 3. Leave-one-out cross validation results (in percentage) Algorithm Majority vote Svm-bool Svm-tf Svm-ntf Svm-tfidf Nb-bool Nb-tf

All messages 67.5 67.1 67.8 72.2 71.9 68.7 68.6

400 messages 50.4 75.1 74.3 75.9 71.7 71.1 71.7

322

B. Yu, M. Chen, and L. Kwok

Because the SVM-bool classifier is easiest to explain, we used the SVM-bool classifier to rank the word presence or absence features based on their discriminative power toward separating “more popular” and “less popular” messages on “400 messages” in the SVM-bool classifier. Table 4 lists the words that indicate the “more popular” and “less popular” messages. Surprisingly, “contest”, “win”, “winner”, and “free” all appear as top indicators of “less popular” messages. It means that fans are not that interested in winning the contests, probably because of the low chance to win. On the other hand, after checking context of the top indicators of “more popular” messages, we found that fans are more interested in trying new dishes, for example the word “try” occurred in 15 messages, 14 of which are in the “more popular” category. Table 4. The most indicative words ranked by the SVM-bool classifier Category Top indicative words more popular try, have, coffee, deal, come, days, october, one, then, like, order, will, who, crunchy, with, million, very, pumpkin, flavors, new, spice, fourthmeal, but, donate, join, early, medium, fall, favorites, served, support Less popular the, week, about, http, win, their, I, watch, from, com, last, free, starbucks, first, can, contest, by, yet, memories, tell, amp, visit, live, almost, family, on, fun, our, meal, photo, video

5 Media Type Analysis Social marketing messages use not only words but also other information types such as links, videos, or photos. In this section we report a statistical analysis to examine the impact of the message types on their popularity. Most of the messages we collected contain a metadata entry called “Media Type” which is either “status”, “link”, “video”, or “photo”. A “status” message has only textual descriptions, usually about the business’s updates, status, information, etc. A “link”, “video”, or “photo” message contains a URL, video, or photo respectively. They may or may not contain text. We conjecture that a restaurant chose a media type when posting a message to Facebook in that Facebook is not likely to manually or automatically assign this metadata for each message. We retrieved the media types from all messages and described their basic statistics in Table 5. We then performed ANOVA test to examine whether there is significant difference between the normalized numbers of “likes” for the four media groups (N=670, five messages did not include media type). The homogeneity test shows that the significance is p